How do I keep two or three AI characters consistent in one video?
Give each character their own reference picture and their own full description in the same prompt, make their designs contrast (hair colour, hairstyle, expression), and fix where each one stands. In group shots the model averages faces, so anything you leave shared or vague drifts toward the middle.
Why faces blend in group shots
A video model renders the whole frame at once. When three people share one prompt, details leak between them: in one of our batches the two friends lost their skin tones and hair and drifted toward the lead character. The more similar the designs, the worse it gets. Three girls with the same long straight hair and the same outfit are close to one face to the model.
How many references you can pass
Enough for one or two pictures per character on most current models. Seedance 2.0 takes up to 9 images per generation[1] and Seedance 2.5 up to 30[2]. Kling 4.0's Omni Reference combines up to 15 assets[3]. Veo 3.1 takes up to three reference images[4], which is one per character for a trio and nothing to spare. Number the references and keep the order fixed: the same character is always the same number.
The recipe we use
- Design for contrast. Our trio is platinum-blonde in a high ponytail, long straight jet-black hair, and shoulder-length dark curls. Different expressions too: haughty, sly smirk, deadpan.
- One reference and one description per character, in the prompt. Restate each one's hair, skin and expression even when the opening frame already shows them. Leaving it to the frame is what made our friends drift.
- Fix positions. Say who stands where: "in the middle", "on her left", "on her right". Repeat it in the opening-frame prompt and the video prompt.
- Give each a signature. A ponytail whip, a body roll, a deadpan pop-and-lock. Behaviour keeps three similar characters readable even when faces soften.
- Same outfit set, written out. Matching kits are fine; write them in full so the model does not invent a different one for each person.
- Avoid list: keep "No change to her face" and add "Keep all three identities exactly."
This is a short excerpt from the reference block of our video Own It, three characters on Seedance 2.0 Fast:
@Image1: The opening frame: the three girls standing far back in the bright tent entrance in a V formation (platinum-blonde Malfoid in the middle, Thea with long straight jet-black hair on her left, Odette with dark curls on her right), their fitted kits and the tent. Open on exactly this composition. @Image2: Thea's face. @Image3: Odette's face. Keep all three identities exactly.
Which camera works best
A locked camera with everyone full body in frame. In our group videos the camera is a fixed object in the scene (a camera low on a bench, a phone propped on goblets) with "the only viewpoint, no cuts, no camera movement". Moving cameras and invented cuts give the model more chances to redraw faces. If you need a close-up, render it as its own clip from a frame of the group shot. More on single-character drift is in why your AI character's face changes.
Where Cast fits
- You want two or three characters in one clip without writing the reference block: Cast takes up to three characters per video and sends each one's picture and description as numbered references in a fixed order.
- You want to try a group video free: the Malfoid & Friends pack (Malfoid, Thea and Odette) is premade and free, and the first 5-second video needs no sign-up.
Where it does not:
- Groups larger than three.
- Group shots of real people or of characters from your own photos: Cast draws characters from text descriptions only.
- Scenes longer than 15 seconds in one clip.
People also ask
- Why do my AI characters start to look alike?
- The model renders the frame as a whole and averages faces that share a prompt. Contrasting designs and a separate description per character reduce it.
- How many characters can I keep consistent in one AI video?
- Three is the most we have rendered together, with one picture each, and it held once each was described separately. The model limits are about reference counts, for example 9 images on Seedance 2.0[1] and 3 on Veo 3.1[4].
- Should I put all characters in one reference image?
- Use an opening frame with everyone in it, plus a separate face reference for each extra character. The opening frame fixes the composition; the separate faces fix identity.
- Which camera angle works best for group shots?
- A fixed, eye-level or slightly low camera with everyone full body in frame and no cuts. We have kept the camera fixed in every group video we published.
Sources
Claims about other companies come from their own pages, read on the date shown. Products and prices change; check the link for the current version.
- ByteDance Seed: Seedance 2.0 official launch (read )
- ByteDance Seed: Introducing Seedance 2.5 (read )
- Kling AI: AI video generator (Kling 4.0) (read )
- Google: Veo on the Gemini API (read )
Related answers
- How do I keep an AI character consistent across videos?
- Why does my AI character's face change between videos?
- What is the best AI video generator for consistent characters?
- Can Seedance make a character dance to my song?
- How do I write a good Seedance prompt?
Part of the Cast answers library (Character consistency).
Updated