What stays the same: character continuity across AI scenes

Generating one beautiful character is easy. Generating the same character twice is where AI storytelling actually begins.

Every model wants to drift: a slightly different jawline here, new freckles there, hair that changes texture between scenes. For a single image nobody notices. For a campaign, a film or an editorial series, drift kills the story. The audience stops believing the person exists.

What works, on a principle level:

• Identity as a checklist, not a mood. Concrete features beat atmosphere. Bone structure, eyes, skin, hair and unmistakable markers. The specific is what survives regeneration; the vague is what drifts.
• One canonical reference. A single neutral portrait becomes the source of truth that everything else is measured against. Resist the temptation to keep rerolling it.
• Change the scene, never the person. Scene prompts modify environment, light and action. The identity block travels unchanged.
• Check every frame. Continuity is quality control, not luck. Each output gets compared against the reference before it ships.

This is the same discipline film productions have always had (script supervision, continuity departments), applied to a medium that regenerates the actor from scratch every single frame. The tools are new. The craft is not.