Audio & Speech

WanSong v1.0: Pure diffusion model generates 5-minute songs with dual stems

No more cascaded pipelines—WanSong outputs vocals and music in one run.

Deep Dive

The Wan Team (Alibaba) has introduced WanSong v1.0, a pure diffusion-based foundation model for music generation that sidesteps the complexity of autoregressive (AR) and cascaded multi-stage pipelines. Where most current systems chain an AR model with a diffusion stage, WanSong directly synthesizes high-fidelity, multilingual songs up to 5 minutes long in a single run. Crucially, it outputs dual stems—vocals and background music—simultaneously, giving producers immediate separation without post-processing.

Beyond raw generation, WanSong incorporates step-distillation for faster inference and provides a clear pathway for fine-tuning and customization, enabling downstream tasks like editing and style adaptation. This design balances efficiency, quality, and controllability for long-form commercial song generation. The technical report details the diffusion architecture and training methodology, positioning WanSong as a simpler yet more powerful alternative to current state-of-the-art systems for both research and practical music production.

Key Points
  • Generates up to 5 minutes of high-fidelity, multilingual audio in a single diffusion pass
  • Outputs separate vocals and background music stems simultaneously, eliminating the need for source separation
  • Supports step-distillation for up to faster inference and is designed for fine-tuning and editing tasks

Why It Matters

A single-model diffusion approach simplifies music AI pipelines and enables rapid customization for commercial creators.

📬 Get the top 10 AI stories daily