YuE2: why a song generator needs an editable score

A music generator can produce a plausible recording while giving you little control over its composition. Asking for a different chorus may also change the melody you liked. YuE2 makes the composition an intermediate object: a readable score that can be inspected and revised before another recording is rendered. That changes what an editing request can mean.

AI Search’s music walkthrough points to the original repository. We used the YuE2 paper, generation guide, and benchmark documentation to separate the method, practical interface, and reported evidence.

A score separates composition from performance

The paper describes a progression from style and lyrics to a symbolic melody-and-chord plan, semantic music tokens, acoustic latents, and decoded audio. Symbolic notation makes notes and harmony explicit. The later stages realize choices such as timbre and expression that a lead sheet does not fully specify.

Consider an illustrative four-bar chorus. You like its melody but want a slower tempo and a different chord under the last bar. An explicit score gives those changes an identifiable target. It also lets you inspect whether the requested change was actually represented before listening to the new performance. The recording still needs review: a correct-looking score does not guarantee a correct audio realization.

The useful distinction is between changing the plan and changing its rendition. A new arrangement can sound different while preserving the tune; a new tune is a different compositional intervention. Keep those requests separate when evaluating an editing system.

Work through the timing

For a regular beat, duration in seconds is 60 / BPM. At 120 beats per minute, a beat lasts 0.5 seconds. At 96 BPM, it lasts 0.625 seconds. If an illustrative phrase has 16 beats, its nominal duration changes from 8 seconds to 10 seconds.

Reducing BPM by 20% therefore increases the duration by 25%, because duration varies inversely with tempo. This is easy to miss when estimating where a chorus should end. The example assumes evenly spaced beats and excludes pickups, pauses, expressive timing, and audio padding; it is not a measurement of a generated song.

For an edit, retain the intended beat count and calculate the new duration before playback. Then listen around the edited boundary. If an extra pause appears, do not mistake the longer waveform for successful tempo control. Check the score and the audio separately.

Read the benchmark as a system comparison

The maintainers’ September 12 comparison uses 192 WildSongBench prompts. SongBench Avg measures an automatic quality aggregate; MuLan measures audio–text alignment. The scatter plot below selects several reported systems and retains both coordinates rather than collapsing them into a universal ranking.

Candidate selection matters. Standard YuE2 selects from two candidates using phoneme error rate; best-of-8 adds a larger selection budget and uses scoring criteria that overlap with evaluation. The published protocols are not matched-compute experiments. A best-of-8 point must not be read as the result of one ordinary pipeline call.

A higher position on this particular quality axis is not proof of human preference, statistical significance, or better control of every requested style. A model can improve one metric and regress on another. The full results include production quality, additional alignment measures, and phoneme error rate. Read those alongside the selected plot when making a deployment decision.

Publisher-reported results

Quality and alignment are separate measurements

Y SongBench AvgX MuLan alignment

Focus the graph and use arrow keys, or tap a point, to inspect values. The complete data table is available below.

• YuE2 (best-of-8)
• YuE2
• Suno v5
• Suno v6
• Mureka 9
• ACE-Step 1.5
• YuE 1

View data table

Graph data table

System	MuLan alignment	SongBench Avg
YuE2 (best-of-8)	0.5051	6.9632
YuE2	0.5068	6.7316
Suno v5	0.5428	6.8721
Suno v6	0.4916	6.5562
Mureka 9	0.4394	6.9377
ACE-Step 1.5	0.4372	6.0118
YuE 1	0.2623	4.9165

Plan a small listening test

Choose three briefs with different requirements: a slow instrumental, a lyric-driven verse, and a song with an explicit chorus transition. Write your own short lyrics so you know exactly what should be sung. Retain the same brief when comparing runs, and listen without seeing which version produced each output.

Use separate questions: Are the words intelligible? Does the style match? Is the melody coherent? Did the requested score edit occur? Was unedited material preserved? Keep the answers separate instead of giving one overall score that hides the failure. If candidate selection is allowed, apply the same budget to every comparison and record how the candidate was chosen.

This is a proposed evaluation, not a Qentrah listening study. Published demonstrations establish examples; a controlled set of your own requests establishes whether the system is useful for your editing task.

Keep the plan and the decoder reproducible

The generation guide exposes stages for planning, semantic generation, synthesis, and decoding. Save the score, effective request, model revisions, and generated artifacts. An unchanged saved plan and an edited composition serve different purposes: preserve the original, then create a new request for the revision.

The guide distinguishes a default listening decoder from the decoder used in the recorded benchmark protocol. Regenerating a song while changing the decoder changes two variables at once. A cleaner decoder comparison uses the same acoustic latents for both decodes. Keep their outputs in separate directories.

For installation, follow the repository quick start or Comfy Org’s native workflow. The documented source starting point uses Linux, Python 3.12, and a BF16-capable NVIDIA GPU with 24 GB VRAM. Check the model and integration licenses independently; Comfy Org labels its package CC-BY-NC-4.0. We did not run GPU inference for this article.

The practical benefit of an explicit composition is inspectability. It gives a musician a place to express intent and a reviewer a place to locate an error. Its value is demonstrated when the score change survives into audio while the rest of the song remains recognizably intact.