MotionBench
Do models understand enough to compose motion graphics with a cohesive narrative and structure?
MotionBench is a benchmark and RL environment for motion graphics: animated type, shapes and layout, written as code and rendered to video. It is built around three questions.
- Given a brief, can a model compose a multi-scene piece whose structure, pacing and transitions hold together as one narrative?
- Given only a rendered video, can a model recover its scene graph, keyframes and easing curves precisely enough to re-render it frame for frame?
- Given two renders of the same brief, does a model prefer the one expert motion designers prefer, as often as the designers agree with each other?
MotionBench is in progress. We will publish its tasks, environment and first results on this page.