FantasiaHelp centerOpen Fantasia

Help center / Guides

Keep a character consistent across AI video shots

Why faces and outfits drift between AI video clips, and what kept them steady in our tests: sharp real frames as references and one locked outfit.

Inside one clip, people hold; between clips, they drift

Unless a line says otherwise, these findings come from our own test clips made on Kling 3.0, a model Fantasia does not offer. The steps for Fantasia use its own models, which we have not tested in exactly the same way.

In four different scenes, the people stayed the same across every cut inside one generation. The trouble started when a second clip had to show the same person: a character who was not in the second clip’s start picture came out as a different person.

So the first rule is to put as much of a moment as you can into one generation; see Multi-shot scene prompts.

Give the model real, sharp frames of the face

When the same man had to appear in a second clip, we tried several kinds of reference and compared the faces side by side:

Sharp real frames, never redrawn by an image model, were the most reliable. A small, soft start picture also made a soft video: upscale it with an upscaler that keeps the face, not an image model that redraws it.

Lock the outfit in every picture

The start picture decides the clothes. When one opening picture put a woman in a beige coat while her portrait wore a mustard one, the video followed the opening picture, and her outfit changed between clips. Once every picture of her, the portrait and each start frame, showed the same mustard coat, checked side by side before generating, the outfit held across both clips. Write the outfit into every prompt that shows the person.

When a face still drifts

Starting a second clip from a wide shot where both faces were tiny, both people drifted even with good references. In one retry, starting the same clip from a medium two-shot with both faces large, both held. We have not pinned down why, but large faces are an easy thing to try.

In one test, a non-human creature did not hold: a baby dragon passed as three real close-up frames still turned into a stiff, toy-like lizard in the later wide shots.

Doing it in Fantasia

On Create Video you choose one of two ways to bring a person in: a start picture under Keyframes, or character references under References. They cannot be combined in one generation, so pick per clip.

  1. Make the person once: a photo, an Influencer portrait, or a picture from Create Image. Save it as a character with a name.
  2. With references: on the References tab, on a model that takes references, type @ and the name in the prompt. Their picture goes with that generation. Describe the opening shot in words.
  3. With a start picture: everyone who must look the same has to be in the picture, in the right outfit.
  4. Either way, describe the same outfit in every prompt.
  5. For the next moment of the same scene, Extend continues from your clip’s last moment.

In Manga, each character’s photo goes with every panel that names them, and a prop the story keeps, such as an umbrella, gets the same written description in every panel, so they are drawn alike from panel to panel.

Questions

Why does my character’s face change between AI video clips?

Each generation only knows what you send it. If the second clip gets a redrawn portrait, a soft clip or no reference at all, the face drifts. Sharp real frames of the face and the same outfit in every picture kept it steady in our tests.

Should I sharpen a face with an image model before using it as a reference?

No. In our tests a “change nothing” redraw changed the face. Use an upscaler that enlarges without redrawing.

Updated November 18, 2026.