AI Can Copy a Performance Now. It Still Can't Direct One.

AI video tools like Runway's Act-One and Act-Two can transfer a human performance onto a generated character—capturing motion and timing from a single-camera phone take—but that process only reproduces a performance someone already gave. It does not generate one from a script, and it cannot evaluate whether the take was the right choice. As filmmaker Frank Houbre frames it, based on his experience producing the award-winning solo animated short Voidborne: the model is a capable camera operator for a performance that already happened.

The three things a director actually does—choosing the emotional read, maintaining performance consistency across a scene, and knowing when to stop iterating—remain entirely outside what any current model can do. Runway's own framing of Act-Two reinforces this: the tool is pitched as a handoff mechanism, not a replacement for directorial judgment. The real gap in AI filmmaking, Houbre argues, is not resolution or frame rate. It's that no model yet has an opinion on which take was the right one—and that opinion is the job.