Costume Changes and Scene States

Quick answer: A model given a reference image copies what is in it — the face, the clothes, the shape of the object — and no amount of text will talk it out of that. So anything that changes the face, the clothes, the form or the structure needs its own reference image, which we call a state. Anything that changes where something is, who is holding it, the light, the weather or what it is covered in is written in words and works fine.

Still from “Rust City Knockout”, an AI short drama generated end-to-end by SceneMixer (6 characters, 1 episode) — Sample: “Rust City Knockout” — Painterly Anime, 1 episode, produced end-to-end by SceneMixer from the script · watch the episodes

States show up as cards next to the main asset in the project view. They are generated from the script read when the story calls for one, and you can add your own — the asset library in this sample project is where they sit.

The dividing line

Ask one question: does the thing itself look different, or does the situation around it look different?

A full change of outfit, a transformation, burns after a fire, a collapsed building, a blade out of its scabbard, a snapped staff — the object itself is different, so it needs an image.

Standing versus sitting, held versus on the table, night versus day, in the rain, covered in dust — these ride on top of whatever the reference shows, and the model applies them from text without trouble. Rain in particular is text, not a state.

One more test: if the two versions could appear in the same frame at the same time, they are not states of one thing — they are two separate assets.

Where states come from

The script read proposes them when the story needs them, within tight caps: at most three looks for a character, two for a location, one for a prop. The caps exist because an unbounded read produces states for everything — a cup that is full, a cup that is empty, a lamp that is lit — and none of those change what the object is.

You can also create, rename and delete states yourself, and choose which episodes each one applies to.

The episode asset library showing generated character sheets, location plates and prop images — This episode's asset library: the characters, locations and props it draws on.

How a state image is made

It is an edit of the base image, not a fresh render. The instruction is four short lines: the state name, what to change, one line holding identity steady, one line for the frame. Nothing else — no style string, no description template.

That brevity is the whole trick. A model given a reference image and a three-thousand-word template reads it as “redraw this reference” and hands back the original almost unchanged. Four lines of edit instruction gets you the change you asked for.

Which image a shot actually uses

Priority runs from most specific to least: a state you picked for this particular segment, then a custom image you set for this episode, then a state assigned to the whole episode, then the base image. Only states that have actually been generated count — if anything is ambiguous it falls back to the base.

Which segments use which state is worked out automatically after the storyboard lands, per segment rather than per episode. Per-episode would be wrong: a flashback look applied to a whole episode puts the character in flashback clothes in every single shot of it.

If the state has no image yet

Nothing breaks. The shot uses the base image and the state is described in words instead — the name and the difference are written into the prompt. For a character that replaces the wardrobe phrase while keeping the facial description; for a location or prop it is added to the material's role line.

Generating the image later gives you the stronger version, and shots generated after that will use it.

Change	State image or text?
Full change of outfit	State image
Transformation, burns, scars	State image
Building collapsed or burnt	State image
Blade drawn, object broken	State image
Standing, sitting, kneeling	Text
Held, dropped, on the table	Text
Night, rain, dust, firelight	Text

FAQ

Why is rain not a state?

Rain is applied on top of whatever the reference image shows, and models handle that from text. A state image is for when the subject itself is a different shape.

Can I have a state for a character standing and one sitting?

You can, but you should not — both could be in the same frame, so they are not two states of one thing. Posture is written into the shot.

Do states cost extra?

They are priced like any other setup image, and on your first project each type has its own small free allowance, separate from the one for base images.

What if I regenerate the storyboard?

Per-segment state picks are recalculated. Episode-level assignments are not affected.

Do I have to generate a state before I can use it?

No. Without an image the state is described in words, which is weaker but works. Generate it when you want the change to be reliable.

Try a costume change

Open a character, add a state, and see the difference between an image and a sentence.

Open a project

Updated September 21, 2026