WanSong v1.0 Technical Report
Binghui Chen, Pandeng Li, Yu Liu, Jingren Zhou
eess.AS, cs.CV
2026-07-16
Alibaba's Wan team generates up to 5-minute multilingual songs end to end with a ~25B pure diffusion model that outputs vocal and BGM stems separately, scoring 7.43% pronunciation error vs 9.86% for SunoV5.5.
Commercial song generators like Suno and Mureka are built on autoregressive (AR) modeling: tokens are produced one after another, so long songs either infer slowly or lose perceptual consistency. Later systems bolt a diffusion stage onto AR pipelines as a refiner, which makes the pipeline a chain of parts. The Wan team at Alibaba asked a different question: can you drop the AR trunk entirely and generate a full 5-minute song end to end with one pure diffusion model, while outputting vocals and accompaniment as separate stems?
Audio is represented by a 1D VAE that compresses 44.1 kHz stereo into continuous latents at a compression ratio of 1024, roughly 43.1 latent frames per second. Ablations show this ratio dominates the quality-efficiency trade-off: dropping from 2048 to 1024 gains 2.86 dB SI-SDR, while keeping the effective ratio at 2048 via patch size 2 actually raises pronunciation error from 19.2% to 20.6% — degradation that a better VAE cannot recover.
Dual-stem tokens are the central design. Vocals and BGM are modeled as independent output streams, but inside each transformer block they are treated as different channels of the same token. The motivation: CFG cannot balance the two. High guidance makes phonemes accurate but suppresses the accompaniment; low guidance restores BGM while vocals blur. Separating the streams lets each get its own guidance, and separation into stems comes free, removing a source-separation step from post-production.
The backbone is a hybrid-MMDit of about 25B parameters: text tokens and audio tokens concatenate into one sequence through a shared transformer, with fully-shared AdaLN carried over from Wan2.1. Training uses flow matching across three pretraining stages (90s, 300s, SFT) on more than 6 million hours of multilingual songs.
RLHF trains one reward model per dimension (musicality, lyric accuracy, prompt alignment), then runs DPO followed by ReFL. DPO supervises every timestep and helps global structure but its pairwise loss is noisy; ReFL only works at low noise, targets fine detail, and is easier to reward-hack. Running them in sequence is meant to be complementary.
Test set: 200 songs across Chinese, English, Japanese and Korean, 10+ genres, about 4 minutes each.
| Model | Musicality (own judge) | PER (lower better) | SongBench Musicality |
| WanSong | 5.49 | 7.43% | 6.17 |
| SunoV5.5 | 4.31 | 9.86% | 5.99 |
| SunoV5 | 4.18 | 22.80% | 6.17 |
| MurekaV7.6 | 3.83 | 12.7% | 6.07 |
| LeVo | 1.69 | 27.11% | 3.96 |
PER leads the runner-up SunoV5.5 by about 2.4 points. On public judges the gaps are small (SongBench Musicality ties SunoV5), and the paper itself notes EvalScore and SongBench were trained on only a few thousand to 11k songs, so it adds its own musicality judge trained on 80k human-annotated songs.
Pure diffusion has been the default in image generation for years; this report brings it to a modality that is minutes long and structurally two-stream, and hits commercial-grade numbers, offering a counterexample to the assumption that long audio requires AR. Direct vocal/BGM stem output is a real engineering convenience for mixing, covers and customization, and the diffusion framework allows step distillation to cut inference cost.
Step distillation appears only in the abstract and conclusion; no step counts or speedup numbers are given anywhere, so the efficiency claim rests on assertion.
The headline 5.49 vs 4.31 musicality gap comes from the authors' own judge, trained on data from the same organization as the model being scored — an obvious bias risk, and the only judge where the gap to SunoV5.5 is large.
Suno and Mureka are commercial black boxes; sampling parameters and version details are uncontrollable, so the comparison's fairness cannot be audited.
The compression ratio was not pushed below 1024; the authors concede compute and speed constraints, so the quality ceiling may not yet be reached.