Synthetic Character Continuity

A face match is the first step

The obvious version of synthetic character continuity is simple: can the model keep the same person across shots?

That matters. Without a stable face, body type, costume, and silhouette, narrative work collapses quickly. Runway Gen-4 is important because it puts character, object, location, and world consistency at the centre of the pitch.

But filmmakers know continuity is not only resemblance. A character can look the same and still feel wrong if the eyeline changes, the emotional state resets, the wardrobe behaves differently, the blocking contradicts the previous shot, or the edit has no reason to cut.

AI video continuity starts as image matching. It becomes cinema only when the scene relationships hold.

Build a character bible before generating

The practical fix is to treat synthetic characters like production characters.

Before generating a scene, define the character bible: face reference, age range, body language, wardrobe, hair, colour palette, emotional register, forbidden variations, and the reasons the character matters inside the story. A prompt alone is too fragile.

This is the point where cinematography, costume, production design, and directing overlap. The character is not only a person-shaped image. They are a set of rules that the camera and edit must respect.

Small teams can do this with a folder and a table. The sophistication is not in the software. It is in the discipline.

Runway's object consistency examples make the continuity problem visible: references have to persist across the world, not only in one frame. Image via Runway Gen-4.

Continuity lives in geography

AI video often treats character continuity as if the character floats in an aesthetic cloud. Real scenes do not work like that.

A character exists in space. They cross a room, sit under a source, look toward another person, pick up an object, move from wide shot to close-up, and carry the emotional pressure of what just happened. If the location geography does not hold, the character cannot hold either.

This is why world consistency matters. The model needs references, but the filmmaker needs a map. Where is the window? Which side is the door? What is the motivated light? What axis are we protecting? Which shot gives the editor permission to cut?

Without geography, continuity becomes costume only.

Performance is the harder continuity

The hardest continuity is performance continuity.

A synthetic character can maintain a face while losing the inner action of the scene. The look is consistent, but the person has forgotten the previous line. That is why AI video still needs human authorship around beats, tension, reaction, and withholding.

OpenAI's Sora material describes models that understand motion, scene continuity, and dynamic clips, and Google positions Veo around increasingly realistic video generation. Those are meaningful advances, but the filmmaker's job remains clear: decide what the performance is doing.

The tool can generate motion. It cannot know whether the character should move.

Review like a script supervisor

The best synthetic continuity reviews should look less like prompt rating and more like script supervision.

Check identity, wardrobe, hair, props, eyelines, screen direction, emotional state, time of day, light direction, gesture, object position, and edit logic. Name failures specifically. Do not write, 'make it more consistent.' Write, 'left hand must stay on the bag, key remains camera right, rain level unchanged, reaction is restrained not surprised.'

That kind of note turns AI video from a novelty into production material.

Continuity is not a model feature. It is a working method.