YuE2 is a music generation model from Multimodal Art Projection (M-A-P), the successor to the original YuE. Given lyrics and a style prompt, it first writes a melody-and-chord plan (a symbolic score) and then renders that plan as a complete song with vocals and accompaniment. Because the score is explicit, a person or an agent can inspect and edit it, including melody, chords, lyrics and style, before the audio is generated.
The architecture is an AR–NAR Mixture-of-Transformers backbone. It predicts the score and semantic tokens autoregressively, generates acoustic latents with flow matching, and decodes them with a VAE into 48 kHz stereo audio. The same checkpoint is used for text-to-music, zero-shot covers (a source recording is transcribed with SheetSage2 and reinterpreted in a new style), and agent-driven editing. The project reports being competitive with Suno v5/v6 on WildSongBench; these are the authors' own benchmark claims.
The code is on GitHub under Apache-2.0, and the weights (YuE2-3B, and a ComfyUI repackaging by Comfy-Org) are released under CC BY-NC 4.0, so commercial use is not permitted. The README lists Linux, Python 3.12 and an NVIDIA GPU with 24 GB VRAM for unquantized use, with experimental low-memory options. YouTube coverage mentions around 4 GB VRAM when quantized. It can also be run through ComfyUI nodes, an agent skill, or a NOIZ-hosted free web demo.
