YuE2 Generate Full Songs In ComfyUI Locally - Installation And Review (Patreon)
Downloads
Content
Video : https://youtu.be/kzTo8pLUKJY
Video For Chinese : https://www.bilibili.com/video/BV1W8eL6qEgC/
Hey everyone. New video is up, and this one is about Yue 2, an open-source music generation model that just beat Suno on benchmarks.
Video : https://youtu.be/kzTo8pLUKJY
You give it a style prompt and lyrics, and it generates a full song with vocals and accompaniment. 48kHz stereo, English, Mandarin, Japanese, Korean. A 3.6-minute song takes about 71 seconds on an RTX 4090.
The interesting part: Yue 2 writes an ABC music notation score before it generates the audio. That means you can edit the chords, change the melody, and regenerate from your edits. Suno and Udio do not give you that. It is like having a sheet of music you can tweak before the performance.
I tested it in ComfyUI with the workflow from the video. The GGUF quantized version runs on 8 to 9 GB VRAM, so even a 4070 or 3080 can get in on this. Supporters, the full run guide is in the supporter post.
Links in the description. Model is on Hugging Face, ComfyUI version included.
— Benji
Resources:
- Yue 2 Model: https://huggingface.co/m-a-p/YuE2-3B
- ComfyUI Pack: https://huggingface.co/Comfy-Org/YuE2
- GGUF (low VRAM): https://huggingface.co/audio-cpp/Yue2-3B-GGUF
- Demo Page: https://map-yue2.github.io/
- GitHub: https://github.com/multimodal-art-projection/YuE
- Paper: https://arxiv.org/abs/2503.08638