Wardrobe and prop continuity across shots
The face holds and the watch changes wrists. A fixed continuity clause to paste into every shot prompt, plus the costume-and-props check to run at assembly.
Lock a character reference, run six shots, and the face will usually hold. Then you cut them together and the mug is in the left hand in shot two and the right hand in shot three, the jacket has picked up a third button somewhere around shot four, and the watch has moved wrists. Nobody watching says "the wardrobe drifted." They say the edit feels off, and they stop watching.
This is a different problem from identity drift and does not have the same fix. Identity is solved by anchoring. Wardrobe and props are solved by writing, then by checking.
Why props drift after the face is locked
The 2026 default for consistency is reference-first: build one strong portrait, then a small sheet of angles and expression variants, then feed that sheet into every downstream shot rather than re-prompting the character from scratch. It works. It also answers exactly one question — who is this person — and nothing about what they are wearing below the collar or holding in their hands.
That neutrality is a requirement of building good references, not an oversight. The hygiene rules that make references fuse cleanly are uniform lighting, a consistent torso-up crop, and a reasonably high resolution; mixed-condition reference sets are what cause fusion errors, where the model averages two looks into a third person who is neither. Adding a coffee cup to half the frames is precisely the inconsistency you were keeping out.
So identity carries forward through the reference, and everything else is carried by the prompt. Anything the prompt phrases differently in shot three than it did in shot two gets re-decided, because a described garment has far more degrees of freedom than a face does. We broke the full taxonomy of "these shots don't match" into four separate failures — identity, wardrobe, set and lighting — in diagnosing drift between shots. This post is the operational fix for the wardrobe one, plus the prop drift that travels alongside it and that the taxonomy does not name separately.
The continuity clause
One block of text. Written once. Pasted into every shot prompt in the sequence character for character. Not paraphrased, not tightened, not "improved" on shot five because you thought of a better word.
Five slots, in this order:
- Garments, colour named as a colour rather than a mood.
- One countable structural detail per garment — buttons, pockets, straps.
- Worn objects with a side — watch on which wrist, bag on which shoulder.
- Carried objects with a hand and a state — which hand, how full, open or closed.
- Hair as a state — tied or loose, and whether strands are out.
A working example:
CONTINUITY: faded navy canvas jacket, three-button front, single chest
pocket on the left; plain white t-shirt underneath; steel watch on the
left wrist; white ceramic mug in the right hand, half full; hair tied
back, no loose strands.
That is roughly forty words and it goes in every shot in the scene, unchanged. When something genuinely has to change — the mug empties, the jacket comes off — write a second clause so the change lands on a cut. A prop that changes on a cut reads as editing; one that changes between two continuous shots reads as a mistake, and it is one.
What belongs in the clause and what does not:
| Put it in | Leave it out |
|---|---|
| Countable structure: buttons, pockets, straps, rings | Drape adjectives: flowing, billowing, crisp |
| Colour as colour: faded navy, oxblood, bone white | Colour as mood: moody blue, warm neutral |
| Side: left wrist, right hand, part on the left | Position in frame: at frame left, foreground |
| Checkable state: half full, unzipped, sleeve rolled | Backstory: her father's watch |
| Hair as a state: tied back, loose to the shoulder | Hair as styling: tousled, effortless |
The organising principle is that continuity errors people notice are countable ones. A viewer will not register that a jacket's drape changed between shots; they will register that it gained a button, because buttons are discrete and the eye counts them without being asked. Spend your clause length on things with integer values and clear sides.
Everything in the right-hand column belongs in the action clause, where it is supposed to vary per shot, or in the cinematography slot, where it describes framing. Mixing them in is how the clause starts mutating between shots, which defeats the point of having one.
Give every prop a home
A described prop with no stated relationship to the body floats. "Holding a mug" leaves the model to pick a hand, a height and a grip, and it picks differently every generation. State the hand, then anchor the height to a body landmark: "white ceramic mug in the right hand at chest height, fingers through the handle."
There is a side benefit. Hands are one of the surfaces that still expose generated footage, and a rigid held object constrains the pose space the model has to solve — a hand wrapped around a cup is a much narrower problem than a hand free in the air. We covered that and the related surfaces in hands, teeth and signage.
The other rule: props enter and leave on cuts. An object that materialises mid-clip is something the model invented, and once it invents one it will happily invent others further down the chain.
Reference the wardrobe instead of only describing it
Text is the weaker of the two levers. The stronger one is showing the model the outfit. Reference-to-video and image-to-video paths both take an image as the binding input, and an outfit captured once in a still is far more specific than any sentence about it. Render one clean full-body still of the character in costume, holding the prop, and anchor the sequence to it — then keep the continuity clause anyway, because the text is what stops the model reinterpreting details the reference happens to occlude.
Reference hygiene applies to wardrobe stills exactly as it does to face references: one lighting condition, one crop convention, no mixing a clean render with a screengrab. The reference-to-video shortlist is a starting point for models, and reference stacking covers feeding several anchors at once.
The costume-and-props check at assembly
Run this on stills, before you watch anything at speed.
- Pull three stills per shot — first frame, middle, last frame — from the rendered footage. The editor can turn video into still photos directly, so you are not screenshotting a player window.
- Lay them out in shot order at the same size. Same size matters; a continuity difference hides easily between two images at different scales.
- Count. Buttons, pockets, straps, visible fastenings, rings. Integers first.
- Check sides. Watch wrist, bag shoulder, which hand holds the prop, which side the hair parts on.
- Check state. Zipped or not, sleeve rolled or down, mug full or empty — and for each change, confirm you scripted it.
- Check the set props. The object on the table behind the subject is in continuity too, and it is the one nobody looks at.
- Then watch the cut at speed.
The stills-first order exists because of the frame rate. At 25 fps a wrong button is on screen for a fraction of a second and playback review will not catch it — but a viewer who scrubs, pauses, or watches twice will. Static comparison finds in thirty seconds what ten playthroughs miss.
One caveat on tooling: the editor's 480p preview pass is free and carries a short per-user cooldown, which makes it the cheap way to check cut rhythm, but 480p is not where you count buttons. Do the continuity pass on full-resolution stills. The final export is charged once regardless of how many clips are on the timeline.
Fix, or re-shoot:
| Symptom | What to do |
|---|---|
| Small prop wrong — mug colour, watch face, label | Mask the region and regenerate just that area with inpainting |
| Garment structure wrong — button count, collar shape | Re-generate the shot; structure is tied to the whole body pose and does not repair locally |
| Prop in the wrong hand | Re-generate; you cannot mirror a region without breaking everything around it |
| Prop drifts within a single clip | Shorten the clip and re-anchor the opening frame rather than re-rolling at the same length |
The last row is the one people resist. A prop that mutates inside one continuous shot means the generation ran past the point where the model was still holding its own state, and a re-roll at the same duration usually reproduces it.
FAQ
Where in the prompt should the continuity clause sit?
Fixed position, never moved. Stability matters more than placement — a clause that migrates from the end of the prompt to the middle is a changed prompt, which is what you were avoiding. After the subject and before the action clause reads naturally and keeps the action last.
How long can the clause get before it starts hurting?
Long enough to cover countable details on visible garments plus worn and carried objects, and no longer. Every word competes with the action and cinematography for attention, and a clause listing fabric weave and stitching crowds out the camera instruction that decides your framing. Past a short paragraph, move the detail into a reference image.
Doesn't a stronger character reference fix wardrobe on its own?
Partly, and less than people expect. A reference locks what is inside its crop under its lighting. Anything below the crop, occluded, or absent from it entirely is still being generated from text. The clause and the reference are complementary, not alternatives.
What about a scene where the wardrobe is supposed to change?
Write two clauses and treat the boundary as a hard cut. Gradual wardrobe change is indistinguishable from drift, so an audience reads your intentional choice as a bug.