The outfit changes between shot 2 and shot 3
Costume drift is identity drift you only catch at assembly. Reuse one wardrobe block verbatim, and know which garment details need a still instead of text.
Shot two looks like her. Shot three looks like her. Cut them together and the jacket has grown a pocket, the stripe on the sleeve has moved, and the red is a different red. Nobody in the review says wardrobe drifted. They say the cut feels off.
Costume drift is identity drift with a delay. Faces get checked first. Garments get checked at assembly, when it is expensive. Consistency collapse splits "these shots don't match" into identity, wardrobe, set and lighting for exactly this reason: the face can hold while the outfit quietly re-rolls. A described garment has far more degrees of freedom than a face, and nothing in a one-line character prompt pins those down.
Two controls close it. A wardrobe block written once and pasted into every shot prompt, character for character. And a clear split between details text can hold (colour as a colour, countable hardware, which side) and details text cannot hold (prints, logos, wear, drape, anything with a pattern). Those latter details need a still.
This is adjacent to wardrobe and prop continuity, the continuity-clause-plus-props checklist. What follows is the other half: which parts of a costume are eligible for a clause, and which are a reference image or they will move.
Costume drift is identity drift on a delay
Character consistency is whether the same person still looks like themselves across separate generations. The glossary already warns that wardrobe, lighting and lens language read as identity too. That sentence is the whole diagnostic.
Review in isolation and you will sign off shot three. The face matches. The model did the expensive thing. The jacket is "a navy jacket" in both prompts, and both jackets are navy. Only a pair of stills, same scale, side by side, shows that shot two is a two-button peak lapel and shot three is a three-button notch. Viewers do not name that. They feel a cut that was never made.
So the review habit has to change. After identity, before you fall in love with a take, check:
- Button count, pocket count, zipper presence.
- Stripe and piping position, not just "striped."
- Jewellery and watches, including which wrist.
- Hem length against a body landmark (hip, mid-thigh, knee).
- Logo or graphic: word count and placement, not "the brand mark."
If any of those disagree, it is not a grading problem. It is a wardrobe that was re-sampled. Treat the shot as failing identity, because that is how it will play.
The delay is also why this gets misfiled as "the model is inconsistent." The model is consistent with an under-specified garment. You asked for a navy jacket twice and received two draws from that distribution.
What text can hold, and what it cannot
Text is good at integers, sides, and colour names. It is bad at surfaces that have a lot of legal variations, all of which match the same words.
| Survives a verbatim sentence | Needs a still |
|---|---|
| Colour as a colour: faded navy, oxblood, bone | Colour as mood: moody blue, warm neutral |
| Countable hardware: three buttons, two chest pockets, one silver zip | Drape: how the fabric sits, crease pattern, "tailored fit" |
| Side and state: watch on left wrist, sleeves rolled to the elbow | Prints, plaids, stripes whose spacing matters |
| Named fibre that changes the silhouette: heavy wool, thin cotton tee | Logos, badges, album art, any glyph |
| Hair as a state: tied back, centre part, no loose strands | Wear: fading map, stains, unique distressing |
A viewer will not register that a jacket's drape changed. They will register that it gained a button, because buttons are discrete, and they will register a graphic that rearranged itself, because graphics are read. Countable hardware, and anything with a pattern or a glyph, are the classes worth spending control on.
Prints are the trap. "Blue and white striped Breton shirt" feels specific and does not pin stripe width, stripe count, or whether the stripes hit the shoulder seam. The model will give you a Breton. It will not give you the same Breton. If the stripe is part of the character, the shirt needs its own still.
Logos are not a clause problem. In-image text is rendered as visual texture, not spelled. A jacket badge that must say the right word is a composite job, or a still of the real garment.
Drape language in the clause is worse than useless. "Crisp," "flowing," "lived-in" re-sample fold patterns every shot. If you need a specific wrinkle, write it as a state with a landmark, or show it.
One wardrobe block, reused verbatim
Write the block once, before shot two exists. Paste it into every prompt in the sequence. Do not paraphrase it on shot five because you thought of a better word. A synonym is a new draw.
WARDROBE: faded navy canvas jacket, three-button
front, single welt pocket on the left chest, no
other pockets; plain white crew-neck t-shirt
underneath; black straight jeans, no wash
pattern; steel watch on the left wrist; white
canvas sneakers, no logo. Hair tied back, centre
part, no loose strands.
That is the garment half of a continuity clause. Props (mug in which hand, bag on which shoulder) belong in the same paste if they appear. The rule here is about mutation: the block is a constant. The action line is the variable.
ACTION: walking left to right, glancing at a
phone in the right hand, overcast street.
When something in the wardrobe actually has to change — jacket off, sleeves rolled, hair down — write a second block and put the change on a cut. A costume change between two shots meant to be continuous reads as a mistake. A costume change on a cut reads as editing.
Register the block where you will actually find it. A named character asset with the wardrobe stills attached is better than a paragraph in a doc. Set the character up once and point later shots at the name. If you also paste the clause, you have two anchors that agree rather than two sources of truth that compete. Do not re-describe the jacket in the action line "just to be safe." Two descriptions of the same garment are how you get a third jacket.
When the garment needs its own still
If a row in the right-hand column of the table above matters to the character, stop trying to name it. Show it.
One clean full-body still, even light, garment occupying enough of the frame to read hardware and print, is the wardrobe reference. A beauty close-up of the face with a sliver of collar is not. You cannot lock a stripe from a collar.
How you feed it:
- Image-to-video for a shot that can start on that still. Frame zero is the outfit. Prompt only the motion. Image-to-video versus reference is the distinction: the still is the first frame, so identity and wardrobe are guaranteed at t=0 and decay from there. Fine for a hold. Weak for a new angle.
- Reference-to-video for a new angle of the same outfit. The wardrobe still is an identity of a garment, not a frame. Pair it with the face sheet so the model does not have to choose which one to honour. A face slot plus a garment slot is the minimum kit the moment prints or hardware have to survive a new framing.
- A dedicated still of the garment on a hanger or a dress form if the on-body shot hides the print. Pattern lock is a still of the pattern, not a still of a person wearing it in three-quarter where half the graphic is around the ribcage.
Do not mix a studio wardrobe still with a frame-grab from shot two. Shot two already contains whatever drift that generation introduced. Chaining wardrobe off the last clip is how a slightly-wrong pocket becomes the new correct pocket. Re-anchor the garment to the master still.
Treat the wardrobe still as a frozen asset: it does not update when you like a take. At assembly, check wardrobe on stills before you watch at speed. Pause two adjacent shots, same scale, scan the list above. If only the colour shifted and the hardware held, you may be looking at lighting drift rather than costume drift.
FAQ
The face is locked and the jacket still changes. Do I need a better face reference?
No. The face reference is doing its job. It answers who. It does not answer what they are wearing below the collar, and a good face sheet is often cropped that way on purpose so the model does not fuse a particular fold of fabric into the identity. Add a wardrobe still, or a stricter clause, depending on whether the drifting detail is a print or a button count.
Can I describe the outfit once in shot one and leave it out later?
You can, and shot three will draw a new outfit. Nothing carries between generations by default. The clause is the part of the prompt that has to be identical every time, including the shots where nothing changed.
What if the drifting detail is a graphic on the shirt?
Do not inpaint letter-level copy across a whole sequence if you can avoid it. Generate with a blank plate or a solid colour in that region and composite the real graphic, or lock the garment with a still of the real shirt. Text-in-image is a texture problem. Wardrobe clauses do not spell.
Does a style reference lock the costume?
No. A style reference locks palette, grade, lens character. The failure signature is exact: every shot looks like the same photographer and a different wardrobe department. If the grade holds and the buttons do not, you solved a style problem you did not have.