Why AI keeps losing your character.

What is character consistency?

Character consistency is the property of a generated sequence in which one character reads as the same person in every shot: the same face, build, hair, and wardrobe, under changing angle, light, and motion.

It is the difference between a striking clip and a film. An audience will forgive a soft frame or an odd cut, and will not forgive a lead whose face rearranges between shots. Consistency is what turns generated footage into a character an audience can follow.

Why do models lose a character?

Because most generation is stateless. A text-to-video pass samples each clip independently, rebuilding the character from the prompt every time, and everything the prompt leaves unsaid gets resampled.

Three properties of the models make it worse. Identity lives in fine detail while the sampler spends most of its capacity on composition and motion, so faces are the first thing to go. Identity is entangled with pose, lens, and lighting in latent space, so changing the angle moves the face along with it. And error compounds across a clip's own frames, so drift grows with duration even inside a single shot.

A written description does not fix this. "A woman in her thirties, dark curls, olive coat" describes a distribution, not a person, and the model draws a fresh sample from it each run.

What actually holds a character?

Seven techniques, in rough order of leverage per hour spent.

1. Lock the character in stills first. Build a reference sheet before any video: front, three-quarter, profile, full body, in two lighting setups. Video inherits every inconsistency the stills carry, so this is the cheapest place to fix them.
2. Condition on an image, not a sentence. Reference-image and image-to-video conditioning constrain identity directly. This single change does more than any amount of prompt rewriting.
3. Train the identity when the shot count justifies it. A LoRA or character embedding on roughly 15 to 40 clean images gives the strongest hold available. It costs hours, so it pays off past about a dozen shots.
4. Simplify what has to stay the same. One silhouette, one wardrobe, one hairstyle, few small props. Every additional detail is another thing that can drift, and prints, logos, and patterns drift first.
5. Fix the seed and freeze the prompt skeleton. Change one variable per attempt. A rewritten prompt is a new character, even when it means the same thing to you.
6. Keep shots short and cut. Drift accumulates with duration. Three three-second shots hold a face better than one nine-second take, and the cut is free.
7. Patch in post. Face restoration, targeted swaps, and hand paint over the frames that drift. Most consistent work you have seen is partly a post fix, and treating that as cheating is how amateurs stay amateur.

What still fails?

Six failures survive every technique above, and any course promising otherwise has not shipped a film.

• Faces at distance and in profile. Fewer identity pixels, weaker hold.
• Hands, contact, and small props. Still the most reliable tell in the medium.
• Wardrobe detail. Buttons, prints, logos, and jewellery are resampled almost every shot.
• Identity under strong expression or fast motion. A wide smile or a hard turn pulls the face off model.
• Two characters in one frame. Features bleed between them, and the effect worsens when they are similar in age or build.
• Continuity of state. Dirt, sweat, wet hair, and injuries should persist across a scene and do not.

How do you test it?

Never from your best shot. Cut the two most distant shots in the sequence back to back, play them on the largest screen you have, and show them to someone who has never seen the character.

Three checks catch most self-deception. Watch at full resolution, because a face that survives a phone will not survive a projector. Watch in the assembled order rather than shot by shot, because drift is only visible against its neighbours. And ask your viewer whether it is the same person, rather than whether it looks good, because those two questions get different answers.

The society settles the model question with measurement instead of opinion: the same character brief and shot list, judged blind, across every major model family, with the harness published alongside the numbers. The first report lands in September.