learnCharacters
Why your AI character changes in wide shots, and how we plan wides around it
A small figure in a wide frame gives the model room to redraw the face. We shoot wides for the outfit and let the medium shot carry the face.
by reeldouble · November 4, 2026 · 6 min read
The problem with a small face
Look at the picture at the top of this page. The woman in the doorway takes up a few percent of the frame. Her suit and her hair read clearly. Her face is a handful of pixels.
That is the shot where AI characters drift most. When a figure is small, the model has very little of the face to match against your references, and each generation redraws it from scratch. Vidu's consistency guide says a small figure in a complex frame gives the model more freedom to reinterpret. Reallusion's manual says a tiny actor in the start frame weakens identity and recommends adding more references. The odd part is that the drift is hard to see at that size. You find it when you cut to the medium shot and the jaw or the hairline does not match.
Most advice for this stops at "test your references in wide and close framing". This page is about planning around the problem instead: what each shot is responsible for, which references to send, and what a wide costs to retake on each engine.
Give each shot one job
We split the work like a film editor would.
The wide sells the place and the outfit. Silhouette, clothing colour, posture, where she is standing. Nobody reads a face at that size, so we do not ask the wide to prove who she is. If the face is half turned, backlit or in shadow, that is fine. It is often better.
The medium sells the identity. Waist up, face clear, the first time the viewer gets a good look. This is the shot that has to match your reference sheet.
The close-up sells the emotion. By this point the viewer already knows who she is. Our page on fixing character drift covers holding the face in tight shots.
A sequence that opens wide, cuts to a medium within a few seconds and only then goes close lets the medium do the identity work. The wide never has to.
How we write the wide
- Put the outfit, hair and build in the first sentence, with the reference token. Those are the traits the wide shows.
- Leave out face detail. Describing eye colour on a figure this small gives the model something to get wrong and nothing to show.
- Keep the light motivated and away from the face: a doorway behind her, a window to the side, a lamp she walks past.
- Give the shot one action. Walking in, turning, stopping. A wide with three actions is three chances to redraw.
- Match the light direction and any haze in the medium that follows, so the cut reads as one place. Martini's multi-shot guide gives the same advice.
Which references to send
The usual guidance is a body reference for wides and a face reference for close-ups. We send both on every shot, because the character set travels with the episode and switching sets between shots is how outfits drift. What changes is how many slots each engine gives you.
| engine | reference slots | what we put in them for a one-character wide |
|---|---|---|
| Kling O3 Pro | 5 elements | the character, plus the location if it repeats |
| Seedance 2.0 | 9 images | full body, three-quarter, face, plus a location plate |
| Seedance 2.5 | 30 images | the full sheet and every angle we have |
| Veo 3.1 Fast | 3 images | full body, face, one three-quarter; render is pinned to 8 s |
If you only have a face portrait, make a full-body image before you shoot wides. A character reference sheet on ReelDouble is 30 credits and gives the model the outfit from more than one side.
What a wide costs to retake
Wides are often silent or under music, and that matters on the engines where silence is cheaper. Here is a 4-second wide with a character reference on each engine, and the 8-second Veo case because Veo pins every reference render to 8 seconds:
| engine | 4 s silent | 4 s voiced | note |
|---|---|---|---|
| Kling O3 Pro | 96 | 120 | 720p only |
| Seedance 2.0 at 720p | 180 | 180 | silence costs the same |
| Seedance 2.5 at 720p | 280 | 280 | silence costs the same |
| Veo 3.1 Fast | 160 | 240 | billed as 8 s with a reference |
So the cheapest wide on our shelf is a silent Kling O3 Pro shot at 96 credits, and a retake is another 96. For an establishing shot that only has to show the outfit and the room, that is where we start.
When to leave the character out of the wide
Sometimes the cleanest fix is no character at all. If the wide is there to show a house, a street or a gym, render it empty and let the character enter in the medium. An empty location on Kling O3 Pro needs no reference and costs the same 24 a second silent. It also gives you a clean plate to reuse as a location reference later, which our same character, different location page uses.
The other option is to keep the figure but turn her away: walking toward a door, looking out a window. A back view carries outfit and hair, which is all a wide needs to show.
A sequence we use
| shot | framing | engine | length | credits |
|---|---|---|---|---|
| 1 | wide, empty hallway | Kling O3 Pro, silent | 4 s | 96 |
| 2 | wide, she steps into the doorway, backlit | Kling O3 Pro, silent | 4 s | 96 |
| 3 | medium, face clear, first line | Kling O3 Pro, voiced | 6 s | 180 |
| 4 | close-up, reaction | Kling O3 Pro, voiced | 4 s | 120 |
Eighteen seconds, 492 credits, and the only shots that have to hold the face are the two where the face is big enough to check. Build it shot by shot in the studio with the same character tagged on all four.
FAQ
Why does my AI character look different in wide shots?
When the character is small in the frame, the model has few pixels of face to match against your references, so it redraws the face more freely each time. Plan wides around outfit and silhouette and let the medium shot carry the face.
Should I use a face reference or a full-body reference for a wide shot?
Send both if the engine has room. The wide mostly shows outfit, hair and build, so a full-body image does most of the work there. On ReelDouble, Kling O3 Pro takes five elements and Seedance 2.0 takes nine images.
How do I stop the face changing between a wide shot and a close-up?
Cut from the wide to a medium shot where the face is clear before you go close, and keep the same references on every shot. Match the light direction across the cut so the faces are lit the same way.
What is the cheapest way to render an establishing shot with an AI character?
On ReelDouble a silent 4-second Kling O3 Pro wide is 96 credits, the cheapest wide with a character on the shelf. An empty establishing shot with no character costs the same and gives you a clean location plate.
Do I need the character in the establishing shot at all?
No. If the wide only shows the place, render it empty and bring the character in on the medium shot. A back view or a backlit figure also works, because a wide only needs to show the outfit and silhouette.


