What each family is good at, and what it leaves open
Sequence prediction
- Good at: local detail that unfolds naturally, one event at a time
- Leaves open: long-range structure over minutes
Representation learning
- Good at: overall structure, key, texture, and style
- Leaves open: choosing the specific next note
Style imitation
- Good at: continuing existing material consistently
- Leaves open: user control beyond the starting fragment
Text guidance
- Good at: letting the user steer with words
- Leaves open: guaranteeing the result is musically coherent
Why combining them is the normal case
A finished song needs all four qualities at once: detail that sounds right moment to moment, a structure that holds over its whole length, a consistent style, and some way for the user to shape the result. Because no single family supplies all four, systems typically run several of them together, letting one handle structure while another handles detail, and letting text guidance sit on top as the user's control. This is why the families are best understood as a set of tools rather than as competing options.
What text guidance adds
Style imitation can only continue what it is given, so the user's influence is limited to choosing the starting material. Text guidance adds a separate channel: the user can describe a mood, a genre, or an instrumentation and have that description shape the output, even when no starting fragment is provided. That is a different kind of control, not a stronger version of the same one.