How Do I Keep Outfits, Props and Injuries Consistent in AI Video?

The one-question test

Continuity in AI video fails in two directions. Either something that needed its own picture was left to words — and the render improvises — or something that words handle fine was frozen into a picture, and now four nearly-identical references fight over every shot. The dividing line is not "how big" the change is; it is what kind of change it is.

A model given a reference image copies what is in it, and no amount of text talks it out of a face, a garment or an object's shape. So anything that changes what the thing is — its identity, clothing, form, structure — needs a reference image of its own. That is what a state is. Anything that changes the situation around it — position, holder, light, weather, what it is covered in — rides on top of the reference and is written in words.

There is a second test for the ambiguous cases: if the two versions could stand in the same frame at the same time, they are not states of one thing — they are two separate assets. A sword in its scabbard and a sword drawn are one prop in two configurations; a sword and a dagger are two props.

How a state gets made

A state image is an edit of the base sheet, not a fresh casting call. The instruction is deliberately short — the state's name, one sentence saying what differs, one line holding the identity steady, one line for the frame — and that brevity is the trick: a three-thousand-word description reads to the model as "redraw this reference" and hands back the original almost unchanged. "Wedding: same face and hair, white gown, no glasses" changes one thing and keeps everything else.

Every state image is a generation at the sheet price: 7 credits per image, with a free allowance for the first project — and writing the change as text before generating costs nothing, which is why learning the state-versus-sentence line saves money rather than costing it. The caps keep the system honest: at most three looks for a character, two for a location, one for a prop. An unbounded read would propose a state for everything — a cup that is full, a cup that is empty — and none of those change what the object is.

Which segments use which state

After the storyboard lands, state assignment is worked out per segment rather than per episode — deliberately. Apply a flashback look to a whole episode and the character wears it in every shot of it, including the ones the storyboard says are present-day. The priority runs from most specific to least: a state picked for this particular segment, then an image set for this episode only, then a state assigned across the episode, then the base sheet.

One more property worth knowing: if a state has no image yet, nothing breaks. The shot uses the base reference and the state is described in words instead. That fallback is why a storyboard can run ahead of your state generation — but it is also why a missing image can quietly soften a look you meant to lock.

Wardrobe arcs, props in two configurations, and injuries

The classic continuity script describes a wardrobe arc — early episodes in one thing, later in another, plus a wedding look. Handed all of it as one description, the render draws a figure wearing half of each. The read splits the arc structurally: the first outfit stays on the base sheet, the rest become separate state images, and part labels like footwear or accessories are recognised as parts of one outfit, not new ones.

Injuries follow the same logic as damage. A bandaged arm, a fresh scar, soot on a face after a fire: each is a new look for a specific stretch of the story, and each is a state selected for exactly the segments where the story has it. Healing works the same way in reverse — the "before the fight" look is the state you switch back to. Props are the same idea with a smaller budget: one state each, and it goes to the configuration problem — the object as carried, drawn, opened, assembled.

Where the continuity labels sit

Continuity labels sit high in the corpus's quality conversations, and they cover the whole body of the problem, not just faces:

Continuity is also what viewers notice when it works. In one thread's comments, the detail that got called out was not the camera or the acting:

I love how even the syringe has a thin wire and is used throghout the effect, like a prop.

A prop that survives the whole effect — parts in the same places, shot after shot — is the difference between an audience reading a tool and an audience reading a mistake. That reading is decided in the props panel, not in post: the reference fixes what the object is, and the words decorate where it goes.

Where this does not apply

• States cost images, so the discipline cuts both ways: an outfit worn once in the background of a crowd scene does not need its own reference. Reserve states for changes the story means to be seen.
• The caps — three looks per character, two per location, one per prop — are not a style guide. If a character genuinely needs more, merge two similar looks into one rather than deleting the change entirely.
• If the change is continuous rather than discrete — a wound slowly healing within one take — no state can help; that is a shot-level choice about what the camera sees.
• A state references a base image, so a wrong base sheet produces wrong states. Fix the sheet first, then regenerate the states that inherit from it.

FAQ

The full mechanics are in the guides on costume changes and scene states and one outfit per character sheet, and the sibling answer on characters that look different in every shot covers the face half of the same problem. The method note describes how these labels are counted; the quoted comment comes from a discussion in r/aivideo.

Pricing·Novel to Video·Guides·Samples·Alternatives·Answers·Support·Privacy·Terms·Legal·Contact·Affiliate program·© 2026 SceneMixer