What the Numbers Actually Show

YuE2's reported numbers are useful only when each one is kept attached to its protocol, sample, metric, and decoder. This chapter reads the WildSongBench comparison, the zero-shot cover results, and the paired editing study as different pieces of evidence rather than as one universal quality score. The practical goal is to state the strongest defensible conclusion from each result and to make the remaining uncertainty visible.

That discipline matters because the headline number is often the easiest part to repeat and the hardest part to interpret. A benchmark lead can describe an observed aggregate without proving statistical superiority. A cover identity score can rise while target-style freedom falls. A local editing measure can support selective control without showing that the old waveform was preserved. YuE2's score-centered design makes these distinctions especially important: the system is valuable partly because it gives users more specific claims to test.

A quality lead is a protocol-specific observation

The September 12, 2026 WildSongBench comparison covers 192 prompts and 17 evaluated settings. Under that comparison, YuE2 in its best-of-8 setting reports a SongBench Average of 6.9632. Standard YuE2 reports 6.7316. The displayed comparison includes Mureka 9 at 6.9377, Suno v5 at 6.8721, and Suno v6 at 6.5562. The strongest safe statement is that YuE2 best-of-8 had the highest observed SongBench Average in this table under the recorded protocol.

That sentence contains several limits. “Best-of-8” is not the same operation as one ordinary pipeline call. Standard YuE2 selects a lower-PER candidate from two generations. Best-of-8 selects from eight candidates using SongBench Musicality, then prompt control, then PER. The chosen output therefore reflects a candidate-selection process as well as the generator. A reader who runs one request once has not reproduced the aggregate setting merely by using the same model name.

The decoder is another boundary. Both YuE2 benchmark settings use YuE2-Vae-legacy, the decoder named by the evaluation protocol. The default YuE2-Vae is intended for generation and listening. That distinction does not make one decoder “real” and the other “artificial”; it means that audio produced under the benchmark protocol and audio produced for ordinary listening have different recorded conditions. A comparison that changes the decoder without changing the latent should be described as a decoder comparison, not as a new model generation.

The table also contains disagreement among metrics. SongBench Average is the mean of seven dimensions, while SongEval is a separate automatic evaluator with its own dimensions. AudioBox production quality, MuLan and AllMusicCaps alignment, prompt control, and PER answer different questions. The repository reports that Suno v6 leads some SongEval and lower-is-better PER outcomes, while YuE2's observed lead is in the SongBench aggregate. This means the result is not a clean ordering from “best” to “worst.” Different systems lead different dimensions.

The gaps should also be read descriptively. The repository explicitly says that the small differences between the highest observed means do not establish statistical significance. The comparison is a documented systems comparison, not a matched-compute experiment that controls every resource and procedure across all evaluated systems. It supports a claim of competitive frontier quality under the stated evaluation, not a universal claim that YuE2 is preferred by listeners or dominates every metric.