Each major family of AI music generation approaches solves a different musical problem, and the problem it solves is the clearest way to tell them apart.
Sequence prediction builds music one event at a time. At each step the system looks at everything generated so far and chooses what is most likely to come next, then repeats. This is well suited to material that unfolds in time, such as a melody line or a drum pattern, because each choice depends on the choices before it.
Representation learning works at a different level. Instead of committing to one next note, it tries to capture the underlying structure and style of a body of musical data, so that the important regularities of that music can be described and reused. This is what allows a system to hold onto a sense of key, texture, or genre across a longer stretch.
Style imitation and continuation takes existing music as its starting point. Given a piece or a fragment, the system extends it by matching its patterns, producing something that sounds like a natural continuation of the same material.
Text-guided generation connects a language description to musical output. A phrase such as a mood, a genre, or an instrumentation request is turned into guidance that shapes what is generated, which is a form of control that style imitation alone cannot offer.
These four jobs are complementary rather than competing. A melody that unfolds convincingly still needs an overall structure, a consistent style, and a way for the user to steer it, and those needs are met by different families working together.