Wan Team's WanSong Generates 5-Minute Songs With Dual Stems

A short technical report from the Wan Team went up on arxiv describing WanSong, a diffusion-based music generation model that outputs full songs and, unusually, keeps the vocal and background music tracks as separate stems in the same generation pass. The paper puts the maximum length at five minutes in a single run and pitches the system as commercial-grade song creation rather than a research toy.

The interesting technical choice is that this is a pure diffusion system, not autoregressive and not a cascaded multi-stage pipeline. Most song generators built to date have leaned on some combination of a language-model backbone with a diffusion decoder, so a claim that a diffusion-only design can reach five minutes with two stems in one shot, and can be sped up further with step-distillation, is what makes the report worth reading. The authors also flag that the model supports fine-tuning and customization for downstream editing tasks, which is what makes the stem output meaningful in practice: an editor gets vocal and instrumental to work with independently, rather than a single mixed track.

Why this matters for anyone building on top of song generators: if the stems really are usable and the model can be fine-tuned, the pipeline for putting AI music into a video edit, a game, or a client project gets a lot less painful than pulling a mixed-down track out of a hosted API and trying to un-mix it after the fact.

The honest caveat is that this is a July 2026 arxiv technical report from Binghui Chen, Pandeng Li, Yu Liu and Jingren Zhou, and the abstract does not publish evaluation metrics, benchmark comparisons, or details on how well the stem separation holds up in the long tail. What the reporting doesn't give you is a listen, a preference score, or a license story, so read the framing as a claim from the team rather than a settled result.

If the stems and the fine-tuning story survive contact with independent testing, the group that benefits most is the mid-tier of music tooling, the people building editors, plugins and workflows on top of a base model, because a diffusion base that hands you separated stems is much easier to build against than a black-box mixed output.