Castaly · Blog · Guides
How to Keep a Character Consistent Across AI Video Clips
Repeated descriptions keep only the broad look. Reference images carry a character into new scenes. Which models accept them, how to word the prompt and what it costs.
A video model starts from nothing on every run. Unless you give it something to hold on to, the woman in your first clip is a different woman in your second. There are three ways to fix that, and they give increasingly tight control: repeat a description, start from a still, or supply reference images.
Method 1: repeat the description
The cheapest method is to reuse the same words. These two Kling 3.0 clips open with an identical description:
A woman in her late twenties with a short dark bob, a small scar above her left eyebrow and a worn olive field jacket …
The bob and the jacket carried over. The scar did not stay put: in the market it sits on her cheek. Repeating words gives you the broad look and little else, which is fine for a background character or a quick test.
If you use this method, give the character one or two strong visual handles, such as a colour or a material, and copy them word for word every time. The prompt guide covers this.
Method 2: start from a still
Make one good image of the character, then use it as the first frame. The opening of the clip now matches exactly, because it is your picture. The drawback is that composition and setting are fixed as well, so a new scene needs a new still.
Stills are cheap. On a flat-price image model, each is $0.0212 on Wan 2.7 Image or $0.0332 on Seedream 5.0 Lite, and Seedream's group mode can return up to 15 related images from one prompt. The image-to-video guide has complete examples.
Method 3: use reference images
A reference image gives the model a subject to keep while it composes a new shot. Here is the method on one character. First, a still from Wan 2.7 Image:

Then HappyHorse 1.1 R2V took it as a reference, with a prompt that names what to keep and moves the event somewhere new:
One continuous take, no cuts. Keep the woman in the crimson satin coat and the peeling carpet ribbons from the reference and move the event outdoors: she walks out of a revolving door onto a wet plaza and the patterned ribbons keep tearing up behind her, now lifting off the paving stones and streaming out across the square in a long coloured wake.
The coat and the carpet pattern carried over into a new location, which is the point of a reference. All together, the still and the clip came to about $0.66.
How to word the prompt
- Say what to keep. Both of our reference prompts state it right after the shot type: Keep the woman in the crimson satin coat…, Keep the glass flowers from the reference exactly as they are…
- Say what changes. Describe the new place and action, as you would for a text prompt.
- Do not describe the character again in detail. The image already carries it, and contradicting words can only confuse the model.
- Make a clean reference. One subject, well lit and clearly visible, with no text.
Which models accept references
| Model | Reference images | Reference videos | 5 s at 1080p |
|---|---|---|---|
seedance-2.5 |
Up to 30 | Up to 10 | $3.54 |
seedance-2.0 |
Up to 9 | Up to 3 | $2.33 |
wan3.0-video |
Up to 10 | Up to 5 | $0.99 |
happyhorse-1.1-r2v |
Up to 9, required | No | $0.64 |
wan2.7-r2v |
Up to 5, combined with videos | Up to 5, combined with images | $0.56 |
kling-3.0-omni |
Up to 7, with elements | 1 | $0.51 |
kling-o1 |
Up to 7, with elements | 1 | $0.51 |
Reference images add nothing to the price on any of these. Reference videos do on some: Kling 3.0 Omni and O1 move to a higher rate, and Seedance adds the video's length to its token count.
For several characters or props in one shot, use a model with a high limit. Kling 3.0 Omni and O1 take up to seven references, and Seedance 2.5 takes up to 30.
Keep the movement too
To make a character perform a specific movement, Kling 3.0 and Kling 2.6 support motion control: one reference image of the character plus one reference video of the movement. The rate on Kling 3.0 does not change.
A workflow for a short series
- Design the character on a cheap image model. Try several versions at about two cents each.
- Save the best one. Anything you generate can go to your Library from Creations and be reused as a reference.
- Test the first scene at a low resolution. Check that the look holds.
- Reuse the same reference and the same "keep" sentence for every scene. Change only the place and the action.
FAQ
What is the best way to keep a character consistent?
Reference images. Repeating a description keeps only the broad look, and a first frame also fixes the composition and setting.
Do reference images cost extra?
No. Reference images do not change the price on any Castaly video model. Reference videos can.
How many references can I use?
Seedance 2.5 takes the most, up to 30 images. Wan 3.0 takes 10, HappyHorse 1.1 R2V and Seedance 2.0 take 9, and Kling 3.0 Omni and O1 take 7.
Can I use a first frame and references together?
On most models they are separate modes. Wan 2.7 R2V is the exception: it allows a first frame alongside its references.
Will exact details such as a scar stay in place?
Not reliably from a description, as the Kling example above shows. A reference gives you a better chance, so test a short clip before you commit to a long one.