Keeping characters and products consistent

The hardest problem in AI video is not making one good shot. It is making the second shot show the same person and the same object as the first.

Why prompts alone fail

A prompt that says “a woman in her thirties with dark hair” describes a category, not a person. Run it four times and you get four different women — all plausible, none the same. The model has no memory between generations.

Adding more adjectives makes this worse, not better: you end up over-specifying the description and under-specifying the identity.

Reference-backed identities

StudioAI solves this by never relying on the name. When you register a character:

1. You supply one to ten photographs, tagged by role — face, full body, side, outfit.
2. A vision model derives a factual description of stable physical traits from those images: face shape, hair, build, complexion, distinguishing features. Wardrobe is kept separate, because it changes.
3. Both the description and the approved photographs travel with every generation that features that character.

The photographs are what does the heavy lifting. The written description keeps the prompt coherent alongside them.

Choosing good reference images

The quality of your references sets the ceiling on consistency.

Do:

• Use well-lit photographs with the subject clearly visible.
• Include at least one straight-on face shot and one full body shot.
• Use images of the same person at roughly the same age and styling.

Avoid:

• Heavy filters, motion blur or extreme angles.
• Group shots where the subject is ambiguous.
• Images where the subject is very small in frame.

The same applies to products. A front shot is the single most valuable reference; add side, detail and packaging shots to pin down proportions and materials.

Continuity between shots

Consistency within a shot is the reference system. Consistency between shots is continuity, and StudioAI handles it separately.

When a scene follows another, the Director carries forward what the previous shot actually established — location, lighting, wardrobe, time of day — and only what it established. If the previous scene never specified lighting, nothing is asserted, because inventing continuity is as damaging as ignoring it.

Where two consecutive scenes share a location and the model supports it, the previous shot’s final frame can seed the next one’s first frame, so the cut reads as continuous.

When it still drifts

• Attach the character to every scene they appear in. A scene with no character reference has nothing to anchor to.
• Check the reference role. A photograph attached as a style reference influences the look; it does not place that person in the shot.
• Watch the reference limit. Models cap how many references they accept. When there are more than the cap, identity references are kept first and style references dropped — but it is better to attach fewer, better references than to rely on that.
• Try a model with stronger reference support. Capabilities differ; the model selector shows which handle image references and how many.