LeVo 2: Stable and Melodious Song Generation via Hierarchical Representation Modeling and Progressive Post-Training

🚀 We introduce LeVo 2 (SongGeneration 2), an open-source foundation model achieving commercial-grade music generation. Through a large-scale expert evaluation (20 professionals, 6 dimensions, 100 songs per model), LeVo 2 is proven to outperform open-source baselines and rival commercial systems in Overall Quality, Melody, Arrangement, Sound Quality, and Structure, while matching many models in controllability assessments.
🧠 LeVo 2 adopts a Hybrid LLM-Diffusion Architecture. The Language Model acts as the "Composer Brain", orchestrating the global musical structure and performance logic (solving how to sing and play), while the Diffusion Model serves as the "High-Fidelity Renderer" to synthesize intricate acoustic details. At its core, we pioneer Hierarchical Representation Modeling to perfectly balance musicality with audio fidelity. Crucially, we utilize a Multi-Stage Progressive Post-Training strategy to explicitly align the LLM with both controllability standards and music aesthetics prior. This alignment process drastically enhances the model's controllability, effectively eliminates lyrical hallucinations, and ensures the generation of stable, structurally coherent, melodious music.
🎧 We are thrilled to release the LeVo-v2-large and LeVo-v2-medium (coming soon) models, which support multilingual lyrics and offer versatile control via text descriptions, audio prompts, and style presets. Designed for accessibility, it can run locally with 22GB VRAM (Large) and as little as 12GB VRAM (Medium).
⚡ For an immediate experience, we also provide the LeVo-v2-fast version on HuggingFace Space! You can now generate a complete song in under one minute, trading a slight loss in musicality for significantly faster generation speed.