A new file format wants to stop AI-generated storyboards from ignoring the script.
Researchers built PACE, a typed specification that sits between a screenplay and the diffusion models used to generate previsualization art. Instead of describing a shot in free text and hoping the model guesses right, PACE writes down exactly who stands where and what the camera does, at whatever level (script, scene, shot or panel) it belongs to. A compiler turns that specification into both the image prompt and an actual 3D scene measured in meters, and a camera solver positions the virtual camera to match the declared framing. On an 11-scene test screenplay called Automatic Drive, every single-subject panel landed within 1.2% of frame width of its intended position; add a second or third subject and one camera angle can't satisfy every mark, so the system reports the miss instead of hiding it.
The more interesting number comes from testing against 204 real director storyboards: shots that hit the declared framing most precisely drew the least of the described action, while looser framing left more room for the model to actually depict what happens in the scene. That's a tradeoff most previsualization demos never admit to: they show off either a pretty picture or accurate blocking, not both at once. Declaring a character's pose explicitly, on a small 30-shot subset, boosted the model's ability to draw the actual action from 58.9% to 74.4% without hurting framing accuracy, which suggests the fix is more structure, not a bigger model.
It's a reminder that generative image and video tools still need someone to do the boring job of geometry: diffusion models are great guessers, bad surveyors.