生成式音訊
YuE2 Bridges Editable Scores and Full-Song Generation, Updated Evaluation Reveals the Cost Difference of Best-of-8
YuE2-3B first generates an ABC melody and chords, then synthesizes 48 kHz songs with a shared-attention AR/NAR architecture, allowing users to modify the composition directly before audio rendering. The WildSongBench update on September 12 ranks it first among 17 systems, but its leading result uses best-of-8 selection and is not directly comparable to the two-candidate setting used by most systems.
YuE2 breaks long-form song generation into observable intermediate representations. Given lyrics and a style, the model first outputs the lead melody and chords in ABC notation, then generates semantic music tokens, produces acoustic latents through flow matching, and finally decodes them with a VAE into 48 kHz stereo audio. This means tasks such as changing the chorus harmony, preserving the vocal melody, or transforming the arrangement no longer have to rely solely on repeated natural-language-guided sampling: users can edit the score first and then re-render it. The [project documentation](https://github.com/multimodal-art-projection/YuE) also exposes staged interfaces including `plan()`, `generate_semantic()`, `synthesize()`, and `decode()`.
The core generator has approximately 3.59 billion parameters and 28 layers, using an AR–NAR Mixture-of-Transformers architecture. The autoregressive and non-autoregressive experts share attention computation while retaining separate normalization, projection, and MLP components. The former handles score and semantic sequences, while the latter handles acoustic generation. Song creation, zero-shot song covering, and editing all use the same checkpoint; only the source of the score differs. For covers, a separate SheetSage2 model can first transcribe the original audio into a melody, after which YuE2 performs it again with new lyrics and a new style.
The [WildSongBench results](https://huggingface.co/datasets/m-a-p/WildSongBench/blob/main/benchmark/results.md), updated on September 12, cover 192 prompts and 17 open and closed systems. YuE2 best-of-8 achieves an average SongBench score of 6.9632, higher than Mureka 9, Suno v5, and Suno v6 in the table; standard two-candidate YuE2 scores 6.7316. This gap also exposes an important limitation: the top score is obtained by generating eight songs and then selecting among them based on automated measures of musicality, controllability, and phoneme error rate. Other systems use inconsistent candidate counts and delivery protocols, and the small differences in mean scores have not been shown to be statistically significant. Nor is this a human-preference evaluation.
Deployment currently requires Python 3.12, an NVIDIA GPU with BF16 support, and approximately 24 GB of VRAM. The code is licensed under Apache 2.0, while the model weights use the commercially restrictive CC BY-NC 4.0 license. The [model card](https://huggingface.co/m-a-p/YuE2-3B) provides weights and reproducibility resources, but the technical report has not yet been released. Key areas to watch next include single-candidate quality, generation latency and memory measurements, and whether the editable symbolic plan genuinely improves the efficiency of human workflows rather than merely raising automated scores.