WanSong Generates 5-Min Stem Songs

Justgototheeffinmoon · reddit · 2026-07-17

The Wan team released a technical report on arXiv detailing WanSong, a diffusion-based music generation model:

Capable of generating complete songs up to 5 minutes long in a single pass.
Simultaneously outputs both vocal and instrumental tracks during the same generation step.
The authors emphasize this is a pure diffusion approach, not an autoregressive or multi-stage cascaded pipeline.
The paper notes that inference can be further accelerated via step distillation.
Supports fine-tuning / customization for downstream editing and remixing.

The post suggests that if the claims of "stem output + fine-tuning" hold up to independent verification, this type of foundational model will greatly benefit teams building music editors, plugins, and workflow tools, as it is much easier to integrate than black-box mixed audio outputs.

Original post →

WanSong Generates 5-Min Stem Songs

More from Multimodal