AI BriefingPickYuE2-3B: Open-source music generation surpasses Suno v5, supports sheet music editing
m-a-p/YuE2-3B
About the project
Generates complete songs with vocals and accompaniment using only lyrics and style prompts. It achieved a SongBench average score of 6.9632, surpassing Suno v5 on WildSongBench, demonstrating the highest performance among open-source models.
Generated tracks are provided as editable scores in ABC notation format. When an AI agent modifies harmony, melody, or lyrics based on user feedback, the model immediately reflects these changes to render a new version. This enables fine-grained musical adjustments, overcoming the 'black box' limitations of existing generative AI.
It also supports cover generation that reinterprets existing audio sources. The melody of the original track can be extracted and arranged in a new style, showing high accuracy on the SHS100K benchmark. It processes 48kHz stereo audio locally on GPUs with 24GB VRAM without quantization.
On an RTX 4090, it generates a 3.6-minute track in approximately 71 seconds. By combining an AR-NAR Mixture-of-Transformers architecture with Flow Matching, it separates score planning and acoustic synthesis, and integrates with the Hugging Face ecosystem to flexibly configure pipelines.
m-a-p/YuE2-3B
The original page has no description.
text-to-audio
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.