YuE2
YuE2 plans a song before it writes one. An AR/NAR mixture-of-transformers backbone writes a symbolic score and semantic tokens, flow matching turns those into acoustic latents, and a VAE decodes them to audio.
It is the only model in Miso that answers at 48 kHz stereo.
What it does in Miso
Section titled “What it does in Miso”One task, generate.yue2, on the create page. The form is not the guided builder that ACE-Step and the others use, because YuE2 takes a style sentence and tagged lyrics directly and those are the only two things it wants.
Vocals
Section titled “Vocals”Full singing in English and Chinese.
Lyrics are required. YuE2 cannot write an instrumental, and an empty lyric is refused outright, which is why the vocal control on its form has no Instrumental option.
Installing it
Section titled “Installing it”YuE2 needs two packages, a model and a decoder, both into Yue2-3B-GGUF/.
Install them one at a time. The five YuE2 packages share four sidecar files, so two installing at once fail.
Miso keeps the decoder out of the studio’s model list so you cannot try to generate with it, and names it as something to install if it is missing.
Options
Section titled “Options”| Option | Does |
|---|---|
lyrics |
The words, with [Verse] and [Chorus] tags. The song follows the structure you give it |
style |
Genre, instruments, voice and production. Required |
cot |
Planning mode: full, melody or off |
abc |
An existing score to sing to. Needs cot=melody or full |
semantic_max_tokens |
Where to stop |
guidance_scale |
How closely it follows the style |
seed |
Same seed and same words give the same song |
There is no duration. YuE2 works the length out from your lyrics. semantic_max_tokens is a stop, not a target, and it is what you raise when a long lyric is getting cut off.
Scores
Section titled “Scores”With planning on, the decoded ABC score comes back as an artifact alongside the audio, and the take gets a Save ABC download.
You can also hand YuE2 a score instead of letting it invent one. This is the most interesting thing it does, and it has enough rules to need its own page. See Scores with YuE2.
Hardware
Section titled “Hardware”About 9 GB of VRAM with the Q8 model and F16 decoder, measured on a 16 GB card.
Licence
Section titled “Licence”The YuE2 weights are CC-BY-NC-4.0, which is non-commercial. Miso itself is Apache 2.0, and the weights are a separate download you choose to install.
If you are making something commercial, use a different family.