LeVo 2: Stable and Melodious Song Generation via Hierarchical Representation Modeling
- Research area: Audio generation
- Authors: Shun Lei, Huaicheng Zhang, Dapeng Wu
- arXiv: 2507.00002
- Pretraining: An automated music aesthetics evaluation framework assigns musicality-tier conditions to large-scale data, providing musicality priors before preference alignment.
- Progressive post-training: SFT, large-scale offline DPO, and closed-loop semi-online DPO successively improve generation quality, controllability, and musicality.
- Modular extension: Track-specific LMs are trained for acoustic refinement while preserving the aligned semantic planner.
Overview
Full-length song generation must preserve coherence and musicality, render detailed vocal and accompaniment acoustics, and follow lyrics and prompts. Existing language model-based systems face a structural trade-off: mixed-token modeling preserves vocal-instrument coordination but obscures track-specific details, whereas dual-track prediction improves acoustics but requires longer sequences and weakens global planning.
The authors present LeVo 2, a hybrid LLM-Diffusion framework for controllable full-length song generation. LeVo 2 formulates this trade-off as hierarchical modeling:
1. LeLM first predicts mixed tokens for semantic planning, then predicts vocal and accompaniment tokens in parallel for track-specific refinement. 2. A diffusion-based Music Codec reconstructs full-length waveforms.
Key Contributions (Extended Version)
The core contribution of this extended version is an aesthetics-guided alignment training scheme:
Results
Expert listening tests and objective evaluations show that LeVo 2 outperforms open-source baselines across six subjective dimensions and approaches leading commercial systems on several listening metrics.
---
*Auto-collected on 2026-07-01.*