Why does the same person keep changing in AI video? Character consistency needs more than a prompt

You generate four shots for a short film. The first looks excellent. The second is acceptable. By the fourth shot, the protagonist appears to be a different person.

The natural response is to make the prompt longer:

25-year-old woman, short black hair, oval face, brown eyes, white shirt...

That improves description consistency, but it does not turn the description into a permanent identity.

Lesson 015 explained the same issue for still images: text describes a region of possible appearances, not a unique identity card.

Video inherits that problem and adds time.

Real productions keep the actor, not only the description

A TV director does not cast a new person in every scene by repeating a list of facial features. The same actor returns to set.

A reference image is closer to that fixed anchor than a text description is.

Current tools can use references to preserve aspects of identity, wardrobe, scene or style. The exact behavior depends on the model, so references improve control rather than guaranteeing pixel-perfect identity.

Build a character master before generating many shots

For an important recurring character, prepare an approved character master set first.

Useful views include:

• front view,
• three-quarter view,
• side view,
• medium portrait,
• full body,
• primary wardrobe,
• neutral expression.

These images become your visual source of truth.

Runway’s reference guidance, for example, recommends clear and naturally lit character references because extreme lighting or obstruction can make the identity harder to interpret.

A keyframe-first workflow is easier to control

Lesson 037 introduced image-to-video. For a multi-shot sequence, a practical pipeline is:

Character reference
→ Generate a still keyframe for each shot
→ Approve identity, clothing and composition
→ Animate each approved frame
→ Edit the clips together

This separates two problems:

1. Does the shot look right?
2. Does the shot move right?

If the still frame already shows the wrong face, there is no reason to spend video-generation budget animating it.

Identity is more than a face

Viewers notice continuity errors in:

• hair color and length,
• body proportions,
• apparent age,
• clothing,
• jewelry,
• makeup,
• props,
• distinctive marks.

A simple continuity sheet can help:

Character: Mia
Hair: shoulder-length black hair
Wardrobe: cream shirt + dark gray trousers
Accessories: small gold hoop on left ear, black watch
Do not change: hair color, earring, watch

The sheet does not force a model to obey perfectly. It gives you an explicit checklist for accepting or rejecting generations.

A seed is not a character ID

Some generation systems expose a seed, the value that influences the random starting state of generation.

A seed can help reproducibility under a fixed setup, but it is not a universal identity token. Change the prompt, model version, reference, aspect ratio or pipeline and the result can still move substantially.

Do not reduce character consistency to:

same seed = same actor

A better mental model is:

stable references
+ explicit character rules
+ controlled shot inputs
+ human selection

Scenes need continuity too

A cafe can drift just as a face can. Three windows become two; a door moves; furniture flips direction.

For recurring locations, create environment references and approved establishing frames. Character continuity and location continuity are separate problems, and both matter to the audience.

One thing to remember

A prompt can describe what a recurring character should look like, but references, approved master images and continuity checks provide the stronger identity anchor. Treat the character like a returning actor, not a fresh generation in every shot.