How to create a consistent character for AI video

Before you start

You need three things: a short written description of the character, an image tool or a camera to make reference images, and a video model that accepts several reference images per request. As of October 2026, Kling 3.0 Omni accepts up to seven reference images, or one reference video of 3 to 10 seconds plus up to four images, according to Kling's own guide. Veo 3.1's reference mode takes up to three images. Expect most of your first hour to go into choosing and rejecting images.

Steps 1 to 3: define, design, expand

1. Write a character card of five lines. Age range and build. Face shape and one distinctive feature, such as a gap in the teeth or thick eyebrows. Hair color, length and parting. The single outfit the character wears in this series. One prop or accessory, if any. The card is for you, not the model. It stops you from accepting images that contradict each other.

2. Make one anchor image. For a fictional character, generate portraits with an image model until one matches the card, then stop. For a real person, take a well-lit front photo with their permission. The anchor is a front view, shoulders up, neutral expression, even light and a plain background. Everything else is built to match it.

3. Expand the anchor into angles. You want a front view, a three-quarter view from each side, a profile, and one full-body shot in the outfit. For a fictional character, use an image model's edit or reference mode with the anchor as input and ask for one new angle at a time. For a real person, shoot the same angles in one session, same clothes, same light. Make more than you need. You will throw most away.

Steps 4 and 5: cut the set and lock it

4. Cut the set down to the images that agree. Lay them side by side and check the details on the card. Is the parting on the same side? Is the scar on the same cheek? Are the eyebrows the same thickness? One image with the wrong parting pulls some clips toward that version. A good final set is three to five images. If only three truly agree, use three.

5. Lock the set. Name it after the character and the version, such as mara-v1, and never edit it in place. If you later want the character in a winter coat, create mara-v2 with its own full set. Mixing versions is the most common way a series drifts over weeks.

Steps 6 and 7: test, then produce

6. Run a test sheet of four short clips. Use the same set and change one thing per clip: a close-up talking to camera, a medium shot walking, a side view in different light, and a wide shot with other people in frame. Three to five seconds each is enough. Put the results next to the anchor. If the face holds in all four, the set works. If it fails in one situation, such as side light, add a reference that covers it. If it fails everywhere, go back to step 4.

7. Produce with scene-only prompts. From now on the prompt never describes the face, hair or outfit, because those come from the references. It describes where the character is, what they do and how the camera moves. For example: "She sits at a cafe window, stirs her coffee, then looks up at the camera. Slow push-in, morning light."

Common mistakes

Describing the character in the prompt as well as in the references. When the two disagree, even slightly, the model blends them.

Using images from different days or with different makeup. To you it is the same person. To the model it is two people averaged together.

Choosing only flattering front portraits. The model then has to invent the profile, and it invents a different one each time.

Fixing a drifting clip with more adjectives. Drift is often random. Regenerate with the same inputs first, and change the references only if the problem repeats.

Checklist and doing it in RefVid

Before you call a character ready: the card exists; the anchor is a clean front view; the set covers front, three-quarters, profile and full body; every image agrees on hair, features and outfit; the set is named and versioned; four test clips held the face.

In RefVid, the locked set becomes a character asset of 1 to 5 images, with the anchor first as the main view. Every clip reuses exactly that set on Kling 3.0 Omni, so step 5 is enforced rather than remembered. Four 3-second test clips cost 504 credits, at 42 credits per second, and all of them sit on the character's page for the side-by-side check.