One can imagine that resolving this will require a separate model to construct a timeline/keyframe series from a text description then interpolate between them.

Fascinating times, either way this is allegedly being open–sourced, so I'm hopeful others will be able to build on top.