How to keep the same AI character when the location changes

Why a new location breaks a character

A character that holds for three shots in one room often comes back different the moment the story moves outside. The face is a little longer, the hair a shade lighter, the jacket has gained a zip. Nothing in the prompt about the character changed. The world did.

Engines rebuild the whole frame every render, face included, and they rebuild it under the light and colour of the scene you describe. Put a character in a dim bar and the model leans into shadow and harder angles; put the same character on a sunny beach and it softens everything. Practitioners on Reddit and in the Seedance guides describe exactly this: dark rooms make faces read more angular, daylight rounds them off. The location is not just background. It is part of what the model uses to draw the face.

The fix is not one trick. It is three habits, and the one most guides leave out is the third: knowing what each move costs, so you budget the retakes before you need them.

Habit 1: split the prompt in two and freeze one half

Every shot prompt we write has two blocks. The character block describes who is on screen and never changes between shots, down to the punctuation. The scene block describes where they are, what they do and how the camera moves, and that is the only part we edit.

A character block for a recurring lead looks like this:

Mara, a woman in her early thirties, shoulder-length dark brown hair with a centre parting, round face, light freckles across the nose, olive green canvas jacket over a cream knit jumper, small silver hoop earrings.

And the scene blocks for three locations:

Scene: a narrow kitchen at night, warm light from a single pendant lamp, Mara leans on the counter and looks toward the window. Camera: slow push in.

Scene: a pine forest at dawn, low mist, cool blue light, Mara walks toward camera along the path. Camera: steady tracking shot backwards.

Scene: a city rooftop at golden hour, wind in her hair, Mara turns to look over her shoulder. Camera: static, medium shot.

Copy the character block, paste it, never retype it. Rewriting "olive green jacket" as "green jacket" in shot five is enough for the model to pick a different green. We keep the block in a text file next to the project and treat it like a costume department. Our descriptor block guide covers how to write one.

Habit 2: attach the same references every time, at the engine's cap

Words carry a character across one room. References carry it across a whole world. The rule is simple: the same images, in the same order, on every shot of the episode, whichever location it is in.

How many you can attach depends on the engine:

engine	references	how the prompt names them	catch
Kling O3 Pro	5 elements	@Element1 to @Element5	up to 15 s per shot
Seedance 2.5	30 images	@Image1 to @Image30	the deepest set we sell
Seedance 2.0 and 2.0 Fast	9 images	@Image1 to @Image9	15 s max
Veo 3.1 Fast	3 images	image 1, image 2, image 3	a reference shot is always 8 s

For a location-heavy episode, the references should show the character under neutral, even light, from the front, three-quarter and side. If your only reference is a moody night portrait, every daytime shot is fighting it. A clean character reference sheet solves that once for the whole series.

On Seedance 2.5, with thirty slots, we also add one plate per location: an empty frame of the kitchen, the forest, the rooftop. The character images stay first in the list and the location plate goes last, so the prompt can say "@Image1 in the setting shown in @Image6". That keeps the location consistent across shots inside it, which is the other half of continuity.

Habit 3: change one thing at a time

When a location move drifts, it is tempting to fix the references, reword the scene and switch engines in the same retake. Then you have no idea which change worked. We change one variable per retake, in this order:

1. Light words first. Add "soft, even light on her face" to the scene block. Most drift on a location change is lighting drift.
2. Then framing. A wide shot leaves the face only a few dozen pixels for the model to get right. Bring the camera closer for the first shot in a new location so the face is established, then go wide.
3. Then the references. If a specific angle keeps failing, add a reference from that angle.
4. Last, the engine. Seedance 2.5 takes the most references of anything we sell, and Kling O3 Pro ties Veo 3.1 Fast for the cheapest voiced second at 30 credits, so a retake there costs least per attempt. Moving a stubborn shot is the final step, not the first.

What a location change costs

Here is the part page 1 never prices. A new location is always a new render. There is no "same character, different background" discount on any engine, because the model draws every pixel again. So the budget question is how many location shots the story needs and how many retakes you allow for them.

One 8-second shot at 720p in a new location:

engine	silent	voiced
Veo 3.1 Fast, with references	160	240
Kling O3 Pro	192	240
Seedance 2.0	360	360
Seedance 2.0 Fast	400	400
Seedance 2.5	560	560

A three-location episode on Kling O3 Pro, two voiced 8-second shots per location, is six shots at 240: 1,440 credits. Allow one retake for the first shot in each new location, because that is where drift shows up, and add 720: 2,160 credits all in. On Seedance 2.5 the same plan is 3,360 before retakes and 5,040 with them. Knowing that up front is what stops a location-heavy story from running out of credits halfway.

When to carry the last frame instead

If the character walks from one place into the next in one continuous movement, through a door or round a corner, you do not need a hard cut. Take the last frame of the outgoing shot and use it as the first frame of the incoming one, with the same character references attached. The engine starts from a face it has already drawn, which removes most of the drift at the join. We use it for doorways, stairwells and car doors.

It does not work for a true jump, kitchen at night to forest at dawn. A first frame in the kitchen and a prompt in the forest gives you a morph, not a cut. For a jump, render a fresh shot with the references and cut in the edit. First and last frame covers the carry in detail.

The order we work in

1. Write the character block once and freeze it.
2. Build a neutral-light reference set at the engine's cap.
3. Plan the shot list by location, first shot in each location framed closer.
4. Budget one retake per new location.
5. Carry the last frame on continuous moves, cut on jumps.

We built the studio around this: characters are saved with their references once, tagged into any scene, and the quote for every shot is on screen before you render it.

FAQ

How do I keep an AI character the same in different locations?

Keep a fixed character description word for word, attach the same reference images to every shot, and only change the scene line. On ReelDouble the saved character carries its references into every scene automatically, so only the location text changes.

Why does my AI character look different when the background changes?

The engine redraws the face under the light and colour of each new scene, so dark or tinted locations reshape it. Adding soft, even light on the face to the scene line and framing the first shot closer fixes most of it.

How many reference images can I use to keep a character consistent?

On ReelDouble, Kling O3 Pro takes 5 elements, Seedance 2.0 and 2.0 Fast take 9 images, Seedance 2.5 takes 30 and Veo 3.1 Fast takes 3, with Veo reference shots always rendering 8 seconds.

Does changing the background of an AI video cost less than a new render?

No. Every engine redraws the whole frame, so a new location is a full new render. An 8-second voiced 720p shot on ReelDouble costs 240 credits on Kling O3 Pro or Veo 3.1 Fast and 560 on Seedance 2.5.

Can I use the last frame of one shot to start the next location?

Yes, when the character moves continuously from one place to the next, such as through a door. For a hard jump between unrelated places, render a fresh shot with the same references and cut between them in the edit.