Direct the body, not the result
What making AVENOR taught me about directing AI video models: the Director's Lock, spatial persistence and why "look sad" doesn't work.
AI video models are astonishing for three seconds and unreliable for thirty.
Making AVENOR, a cinematic universe made with generative video, I kept losing shots to the same handful of failures. A character turns and speaks into the lens. A cup disappears between two frames. Someone slides across the floor without walking. A coat changes colour mid-scene. Any one of them breaks the spell.
Better adjectives didn't fix it. Rules did.
The Director's Lock
Every AVENOR generation now carries a lock: ten non-negotiable rules handed to the model with the shot. There are two versions: a full lock that opens single-shot prompts, and a short lock appended to the end of every multi-shot prompt. One of them is always present. No generation ships without one.
A few of the rules, paraphrased:
- Eyelines and address. Characters look at and speak to the person they're talking to. Nobody addresses the camera unless directed. Listeners visibly listen.
- Spatial persistence. Everyone's position, pose and distance persists until the direction changes it. Nobody drifts, teleports, enters or vanishes.
- Prop and costume persistence. Objects stay in the hand that holds them. Clothes, hair and jewellery stay identical from first frame to last.
Every rule exists because of a specific failed generation.
Results versus instructions
The deeper lesson came from performance. Human directors give result direction, like "play it sadder" or "more intense". Models respond to that with stock expressions.
What works is describing the body: posture, breath, where the weight sits, and the order in which facial muscles move. AVENOR's performance codex has a translation table for exactly this, from the result you want to the physical instruction that produces it.
The camera has a body too
The same applies to the camera. Where does it live relative to the character? When does it commit to a move? When should it not embody a point of view at all? Writing those down as doctrine made the films feel directed rather than generated.
Why this matters beyond film
Every system I build with AI follows the same pattern: the model is capable, but capability without constraints is noise. The work is writing the constraints (the lock) that turn a capable model into a reliable collaborator.