Guide

How to keep the same character in every AI video shot

Your character looks great in shot one, then shows up in shot two with a new nose and a different jacket. Here's the simple routine that fixes most of it, plus what the retries actually cost.

Updated Oct 20, 2026 · about a 6-minute read

The short answer

Write one fixed character description and paste it word for word into every prompt.
Make a small reference sheet (front, side, back) as still images first.
Attach those same images to every shot. Don't rely on text alone.
Change only the action, camera and place between shots.
Draft cheap, check each take against the sheet, then render the keeper at full quality.

Why the face keeps changing

Each generation starts fresh. The model doesn't remember the person it drew last time, so every prompt is a new guess at who "a woman with red hair" is. Small wording changes make it worse: "jacket" in one prompt and "coat" in the next is enough to get a different outfit.

Seeds don't save you either. Google's own Veo docs say the seed parameter "doesn't guarantee determinism, but slightly improves it." What does help is giving the model the same visual source every time.

Step 1: Write a fixed character description

Pick the traits that make your character recognisable and lock them: age, skin tone, face shape, eyes, hair (style, length, colour, parting), one or two distinctive marks, the exact outfit and accessories, and the overall style.

Then treat that paragraph like a logo. Copy and paste it; never retype it. The free Character Lock tool builds this paragraph for you in about a minute.

Step 2: Make a reference sheet before any video

Generate the character as still images first. It's cheaper than video and it's the easiest place to change your mind. You want:

Front: facing camera, eye level, neutral pose.
Side: a clean profile, same pose and same light.
Back: so hair and outfit hold when they turn away.

Use a plain grey background and soft, even light so no mood gets baked in. Pick the best of each, and that's your sheet. One approved version only. If the face on the sheet is wrong, fix the sheet, not the shot.

Step 3: Attach the same references to every shot

This matters more than anything else here. All four models below can take images as input, but in different ways (see the model notes further down). Use the same files every time. Don't swap in a frame from a take that already drifted, or the drift becomes the new normal.

Step 4: Change only what the shot needs

Build each prompt in the same order: fixed description first, then the one action, then camera and setting. Keep the lighting words the same unless the scene truly changes. One action per shot. Fast head turns and extreme angles you didn't include on the sheet are where faces tend to slip.

Example prompt

Fixed description (identical in every shot), then the shot. This is the default character from the Character Lock tool.

Shot 3 of 10

Mira: a late 20s woman, warm olive skin, Mediterranean features, heart-shaped face, hazel almond eyes, shoulder-length wavy copper-red hair, side part, distinctive: small mole above left lip, freckles across nose, wearing mustard-yellow corduroy jacket over white t-shirt, dark jeans, white sneakers, thin gold hoop earrings. Style: photoreal, cinematic. Same face, same hair, same outfit in every shot. Shot: Mira (the woman in the reference images) sits at a small cafe table by the window and looks up from her phone with a slight smile. Medium close-up, eye level, slow push-in. Soft natural daylight.

For shot 4, keep the first paragraph exactly as is and change only the "Shot:" line.

Step 5: Draft cheap, check, then render

Test the motion and framing at the lowest resolution or the cheaper model tier, then re-run only the takes you like at full quality. Before you keep a take, put it next to your sheet and check face shape, eye colour, hair parting, the marks and the outfit colours. Look at the last second of the clip too; drift often shows up at the end.

Keep a simple tally of how many tries each kept shot took. That number is your real cost.

Model notes (from the official docs)

Seedance 2.0 (BytePlus ModelArk API). Image references aren't priced separately in the official token formula, so adding your sheet costs little. The official worked examples come to about $0.07/s at 480p and $0.15/s at 720p, so drafting at 480p is roughly half price. Video references cost more the longer they are. Failed generations (for example, blocked by moderation) aren't charged, but takes you reject are. Seedance notes and prices

Kling 3.0 (Kling AI API). $0.084/s silent or $0.126/s with native audio at 720p, and $0.112 or $0.168 at 1080p. Turn audio off while you test and add it only on the final take. The official API is prepaid, with packages from $700; resellers list the same rates pay-as-you-go. Kling notes and prices

Veo 3.1 (Google Cloud; Gemini API until Oct 22). Accepts up to three reference images "of a single person, character, or product" to preserve their look, and can also pin the first and last frames. Using reference images forces 8-second clips, and Lite doesn't support them. Standard is $0.40/s with audio ($0.20/s silent on Google Cloud), Fast is $0.10–0.12/s with audio. Heads-up: the Gemini API's Veo 3.1 preview models shut down on Oct 22, 2026. Veo 3.1 carries on as a GA model on Google Cloud, and on the Gemini API Google points to Gemini Omni 1.1 Flash (about $0.10/s at 720p). Veo notes and prices

Grok Imagine Video 1.5 (xAI API). Takes up to 14 reference images at $0.01 each, so your whole sheet plus a close-up costs a few cents. Video is $0.08/s at 480p, $0.14/s at 720p and $0.25/s at 1080p. The 1.5 Lite model ($0.02–0.03/s at 480p–720p) is handy for cheap drafts. Grok Imagine notes and prices

Google Vids (Gemini Omni 1.1 Flash). Not priced per second: clips come out of a monthly allowance on your Google plan, from 6 clips on a free personal account to 500 seconds on Business Standard or Google AI Pro. You can add up to 7 reference images as "Ingredients", extend a scene, and set the clip length (3–10 s). Retries still count, just in seconds instead of dollars. Google Vids notes and limits

Prices are API list prices checked Oct 20, 2026. App subscriptions and resellers price differently, and these change often. Sources: BytePlus, Kling, Gemini pricing, Veo docs, Google Cloud pricing, Vids limits, xAI, xAI pricing.

What retries really cost

A consistent character usually means throwing some takes away, and the price list only shows the cost of one take. Here's a simple example: a 5-second shot at 720p, where it takes 4 tries to get one you keep.

Model (720p)	Per try	Per kept shot
Kling 3.0, audio off	$0.42	$1.68
Kling 3.0, audio on	$0.63	$2.52
Grok Imagine 1.5 + 4 refs	$0.74	$2.96
Seedance 2.0	$0.75	$3.00
Veo 3.1 Standard (8 s, refs)	$3.20	$12.80

Three of those four tries are money spent on takes you don't use. Ten shots like that on Veo 3.1 Standard comes to $128 before audio fixes or edits. That's why drafting cheap and locking the character first pays off.

Plug in your own model, length and retry count in the kept-shot cost calculator.

Your checklist

One fixed description, pasted, never retyped
Front, side and back reference images, one approved version
Same reference files attached to every shot
Only the action, camera and place change
Same lighting words, aspect ratio and resolution across the scene
Draft cheap, check against the sheet, render the keeper
Count your tries so you know your real cost

Build your character →Price a shot