learnCharacters
Same AI character, different expression: which words stay frozen and which change
How to change an AI character's expression shot to shot without losing the face: the frozen identity words, the per-shot expression line, and what a broken face costs.
by reeldouble · November 7, 2026 · 6 min read
The advice stops one step short
The pages that rank for this question agree on the basics. Runway says to keep lighting and angle steady so the expression is the only thing that changes. Renderforest says to add an emotional expression to the character sheet. Magic Hour says to repeat the same character description at the start of every prompt. All three are right.
None of them tell you which words in that description are allowed to change. If the description says "a cheerful young woman with a bright smile", every angry shot is fighting your own prompt. This page splits the prompt in two, ranks expressions by how often they break a face, and prices the retake when one does.
Split the prompt in two
We write every character prompt as two parts.
The frozen block. Identity only, written once, pasted unchanged into every shot. Face shape, eye colour, hair (colour, length, how it falls), skin, age, build, outfit, and the art style. No mood words. Our descriptor block guide covers how to write it.
The shot line. Everything that is allowed to change: the expression, where the eyes look, the head position, the action, the camera.
The rule that does most of the work: an emotion word never goes in the frozen block. "Cheerful", "shy", "grumpy", "smiling" all read as a permanent state, and the engine will drag every shot back toward it. Personality belongs in the script. The block describes a face at rest.
Here is a frozen block and three shot lines for the same character:
Frozen: a 3D animated young woman, early twenties, oval face, large brown eyes, messy shoulder-length brown hair with a side part, light skin with faint freckles, grey zip-up top.
Shot 1: close-up, she smiles with her mouth closed, eyes soft, looking at the camera. Shot 2: close-up, brows pulled down and together, lips pressed tight, looking down at the pan. Shot 3: medium shot, eyebrows raised, mouth slightly open, leaning back from the stove.
Describe the face, not the feeling
Shot 2 never says "angry". It describes what the face does: brows down and together, lips pressed. Adobe's community forum has a thread where users asked for angry or yelling characters and got smiling ones back. With a feeling word, the engine has to guess what the face does. We name the brows, the lips and the eyes instead.
The words we reach for:
| expression | what we write |
|---|---|
| angry | brows pulled down and together, lips pressed tight, jaw set |
| sad | inner brows raised, eyes lowered, corners of the mouth turned down |
| surprised | eyebrows raised high, eyes wide, mouth slightly open |
| happy | closed-mouth smile, cheeks lifted, eyes slightly narrowed |
| worried | brows raised and pinched together, lips parted, eyes moving |
| disgusted | nose wrinkled, upper lip raised, head pulled back |
Which expressions break faces
Not every expression carries the same risk. This is our working order, and it agrees with what the guides and forums report: a vendor guide on distorted faces found a wide-open toothy grin bent the face where a closed-lip smile did not, and a Reddit thread on Kling and Seedance names head turns as the biggest trigger for face warping.
| risk | expressions |
|---|---|
| low | neutral, small closed-mouth smile, frown, raised eyebrow, eyes looking away |
| medium | wide smile, crying with the mouth closed, surprise with the mouth slightly open |
| high | laughing with the mouth open, screaming, shouting, yawning, sobbing |
| highest | any high-risk expression while the head turns |
Most of the high-risk row has one thing in common: an open mouth, which reshapes the whole lower face. When the story needs a scream, we keep the head still, keep the camera still, and keep the scream short.
Keep the peak short
A big expression held for eight seconds gives the engine eight seconds to drift. Two habits help.
Build to the peak. Start the shot neutral and end it on the expression: "she stares at the pan, then her face crumples into a frown". The likeness is set on a calm face first.
Cut on the peak. Put the laugh in a 4-second shot and cut away to the reaction. Shorter clips hold better, which Magic Hour's guide says too.
If a shot keeps bending the face, add one panel to the reference sheet showing the character with that expression. A sheet is 30 credits, and a single extra panel of the hard expression gives the engine something to match instead of something to invent.
What a broken face costs
A face that bends on the peak is a full retake. There is no way to re-render only the mouth, so the engine bills the whole shot again. Here is one 6-second close-up with the character attached, per engine:
| engine | one 6 s close-up | one retake on top |
|---|---|---|
| Kling O3 Pro, voiced | 180 | 360 total |
| Kling O3 Pro, silent | 144 | 288 total |
| Veo 3.1 Fast, voiced, with references | 240 (billed as 8 s) | 480 total |
| Seedance 2.0, 480p | 132 | 264 total |
| Seedance 2.0, 720p | 270 | 540 total |
| Seedance 2.5, 720p | 420 | 840 total |
So we test a risky expression before the real shot: 4 seconds on Seedance 2.0 at 480p is 88 credits, about a fifth of a Seedance 2.5 retake. If the face holds through the test, we render the real shot on the engine we finish on.
Our order of work
- Write the frozen block with no emotion words.
- Mark each shot's expression as low, medium or high risk.
- For high-risk shots: still head, still camera, 4 seconds, build to the peak.
- Test the riskiest expression first, at 88 credits.
- If it bends twice, add that expression to the sheet and test again.
In the studio the character's sheet and descriptor are saved on the character, so the frozen block goes into every shot the same way and only the shot line is yours to change. For the voice side of the same problem, see voice emotion prompts.
FAQ
How do I change an AI character's expression without changing the face?
Keep a fixed identity description with no mood words in every prompt, and put the expression in a separate shot line. Describe the face physically, such as brows down and lips pressed, and keep lighting and camera angle steady so the expression is the only change.
Why does my AI character's face change when they laugh or scream?
An open mouth reshapes the lower face, and a head turn at the same time makes it worse. Keep the head and camera still, start the shot neutral, and keep the big expression to about 4 seconds.
Should I add expressions to my AI character reference sheet?
Add one panel for any expression that keeps breaking the face. On ReelDouble a reference sheet costs 30 credits, which is less than one retake of a 6-second close-up on any engine we run.
What words should never go in an AI character description?
Emotion and personality words such as cheerful, grumpy, shy or smiling. They read as a permanent state and pull every shot toward that mood. Keep the description to face, hair, skin, age, build, outfit and art style.
How much does it cost to retake an AI video shot with a bad face?
A retake bills the full shot again. On ReelDouble a 6-second close-up is 180 credits on Kling O3 Pro voiced and 420 on Seedance 2.5 at 720p, so we test risky expressions first at 88 credits on Seedance 2.0 at 480p.
