WanSong v1.0 Technical Report
Summary
WanSong v1.0, released on July 16, 2026, is a pure diffusion-based music generation foundation model designed for long-form, commercial-grade songs. Unlike traditional autoregressive or cascaded AR+diffusion systems, WanSong directly generates high-fidelity, multilingual songs up to 5 minutes, outputting dual stems (vocals and background music) in a single run. It employs a 25B parameter hybrid-MMDit architecture, conditioned by an LLM text captioner, and a novel dual-stem token scheme to prevent vocal-BGM interference. The model, trained on over 6 million hours of multilingual song data and a VAE with a 1024 compression ratio, also integrates reinforcement learning from human feedback to align with musicality, lyric accuracy, and prompt alignment preferences. This approach enables faster inference via step-distillation and simplifies post-production editing.
Key takeaway
For AI Scientists and Machine Learning Engineers developing commercial-grade music generation systems, WanSong v1.0 demonstrates a compelling shift from autoregressive to pure diffusion models. You should consider adopting a dual-stem token scheme and lower VAE compression ratios, like 1024x, to achieve superior audio fidelity and enable direct vocal/BGM separation. This approach simplifies post-production workflows and enhances controllability, offering a practical pathway to generate high-quality, long-duration, multilingual songs without complex multi-stage pipelines.
Key insights
Pure diffusion with dual-stem modeling enables high-fidelity, long-form music generation and separated vocal/BGM outputs.
Principles
- Pure diffusion models can simplify and enhance end-to-end music generation.
- Separate vocal and BGM modeling prevents cross-interference for better quality.
- Lower VAE compression ratios (e.g., 1024x) are vital for fine audio detail.
Method
WanSong employs a hybrid-MMDit backbone with an LLM text captioner, concatenating text and dual-stem audio tokens. It trains a 1-D VAE (1024x compression) and applies RLHF using DPO and ReFL for human preference alignment.
In practice
- Implement dual-stem token schemes to resolve vocal-BGM interference.
- Utilize RLHF with DPO and ReFL for human preference alignment.
- Prioritize lower VAE compression ratios for high-fidelity audio.
Topics
- Music Generation
- Diffusion Models
- Dual-stem Modeling
- Variational Autoencoder
- RLHF
- Long-form Audio
Best for: Research Scientist, AI Scientist, Machine Learning Engineer
Related on AIssential
See Counsel's argued verdicts on the open AI decisions leaders are weighing →
Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.