Instead of describing rhythm or tempo in words, the actual audio track is embedded directly into the generation prompt as an input, with an instruction that the character's movement (e.g., a dance) should follow its beat. This sidesteps the unreliability of text-described timing — the model has the literal waveform to lock onto rather than an approximation like "upbeat" or "fast-paced."
Pairs with move-by-move-choreography-scripting-style explicit move sequencing: the named moves give the what, the embedded track gives the when. Also complements Blocking Lock via Position Reference (a schematic/map fixes where) as another case of feeding the model a literal reference artifact instead of relying on text to encode something text is bad at conveying — timing, in this case.
Apply: When a scene requires movement synced to music, attach the actual audio file as a prompt input and instruct the model to time the action to its beat, rather than writing a text description of the rhythm.