As a film director, what I ask from my actors is consistency.

They’re trained to repeat the same sequence of movements from take to take, from shot to shot, so the material can be cut together. Continuity is what allows cinema to exist as a constructed time.

When I work with AI, I start with a still image. I look for the right formula, the right words to describe the scene I have in mind.

When I’m satisfied with the image, I animate it. I test movements, select what feels right, and integrate it into the edit.

But my AI actor is not an actor.

She cannot be consistent. Every take is different.

Because what the AI produces is not the recording of an action performed by a body in front of a camera. It’s a statistical guess.

An approximation of what someone like this: this age, this body, this gender, this outfit, under these conditions, might do.

Each movement is inferred from prior data. Not remembered. Not repeated.

No two takes match, because there is no original gesture to return to.

For this new series, I decided to stop correcting that instability and instead work with it.

I start from the exact same still image, generate multiple iterations, and overlay them rather than selecting one “best” take.

There is no attempt to simulate continuity. The movement accumulates instead of resolving.

I place the figure against a black background to remove any narrative distraction. What remains is the body as a field of possible motions.

It’s striking to see how many directions the same body can take when movement is no longer anchored to a recorded action, but to probability.

What we’re watching isn’t performance in the traditional sense, but the visible output of a system trying to decide what movement is plausible.

What interests me here is letting the statistical nature of the gesture stay visible.

Letting inconsistency accumulate instead of being corrected.

The image doesn’t settle into a single action; it hovers between many plausible ones.

What appears on screen is a synthetic body inferred from prior behaviors, replayed as motion without memory.

The score is built entirely from textures, noise, and layered sound design in Suno.