Checkpoint: Three Layers of Control

YuE2 is easiest to judge when its score, generator, and decoder are treated as three connected layers rather than as one mysterious source of behavior. This checkpoint brings those layers together and uses them to answer a practical question: if a user changes something, where did the change occur, and what evidence could show that it worked? The answer lets readers distinguish a score-level intervention from a generation choice or a decoder comparison before applying the system to covers and edits.

The compact model is simple enough to remember but precise enough to prevent category errors. The score is where musical intent is authored or supplied. The AR–NAR generator is where the system infers semantic sequences and acoustic latents from that condition. The VAE and evaluation process are where the generated state becomes a listening result and is judged. Each layer can influence the next, but none should be used as evidence for every claim about the others.

The score, generator, and decoder are different kinds of control

At the score layer, a person, an agent, YuE2's planner, or an audio-to-score system supplies a symbolic composition. The score can be generated from lyrics and style, transcribed from a recording, or copied and edited. Full mode places melody and chords in that plan; melody mode retains a melody condition while leaving more accompaniment freedom; off mode removes the inspectable symbolic plan. This layer answers: What musical structure has been stated as the condition?

The score layer is the most direct place for authorship because it is the most legible. A user can say that harmony should change while a melody should remain, or that a melody should be retained while accompaniment adapts. The system can validate the encoded plan and preserve its identity through the exact prefix. But a score is not the entire performance. It states selected structural decisions and leaves acoustic realization open.

At the generation layer, YuE2 uses its AR–NAR Mixture-of-Transformers to infer what the stated composition becomes in model-facing and acoustic representations. The AR path extends symbolic and semantic sequences causally. The NAR path refines continuous acoustic latents with timestep, position, prefix context, and flow-matching velocity. This layer answers: How does the model realize the condition?

The generation layer is powerful but less directly inspectable. Semantic tokens and latent values carry information that the public interface can save and reuse, yet they are not a second human-readable score. They are the generator's internal working states. A fixed score can therefore lead to different performances under different conditions, and a successful semantic or latent result does not by itself show that every requested note was heard exactly as intended.

At the decoder and evaluation layer, a VAE converts the acoustic latents into stereo 48 kHz audio. The default listening decoder and the benchmark decoder have distinct documented roles. The same latents can be decoded again to compare decoder behavior, but that operation does not generate new music. Listening, symbolic comparison, ASR, alignment, and automatic benchmark metrics then answer different questions about the result. This layer answers: What became audible, and what can be concluded from the measurement?

The boundaries are not merely conceptual. They are enforced by different artifacts and transitions. The symbolic plan includes the score and exact encoded condition.