YuE2 prompting guide: styles, lyrics and section tags

The YuE2 authors describe the split plainly: genre, instruments, vocal character, language and intended tempo go in the style; section tags and the actual words go in the lyrics. Production notes placed in the lyrics risk being sung, so keep them in the style.

There is no fixed list of YuE2 style tags; the field is free text. The upstream examples are short runs of comma-separated descriptors, most important first. A dependable order is:

Language, genre, vocal character, tempo, lead instruments, mood or production texture

• Name the language. The model card lists Chinese and English; the ComfyUI guide also lists Japanese among the singing languages.
• Pick one genre and, at most, one modifier. “Indie folk” and “hyperpop” in the same brief pull against each other.
• Give the vocal a character, not just a gender: “breathy alto”, “gritty male lead”, “close-harmony duet”.
• Three or four instruments are enough; name the lead instrument first.
• Keep it short. In ComfyUI a long prompt reduces the room left for the song itself.

These briefs are starting points written for this guide, not presets built into the model. Change one element at a time so you can hear what each change does.

• Indie folk

  English, indie folk, tender male vocal, 84 BPM, fingerpicked acoustic guitar, brushed drums, upright bass, warm and close.

• Synth-pop

  English, glossy synth-pop, bright female lead, 118 BPM, analog arpeggios, punchy bass, gated snare, late-night energy.

• Mandopop ballad

  Mandarin, pop ballad, clear female vocal, 72 BPM, piano, soft strings, light percussion, restrained first verse and a wide chorus.

• Jazz-funk

  English, jazz-funk, smooth male lead, 104 BPM, electric piano, slap bass, tight drums, horn stabs, dry live-room mix.

• Alt rock

  English, alternative rock, gritty male vocal, 132 BPM, overdriven guitars, driving bass, live drums, loud chorus.

• Lo-fi hip hop

  English, lo-fi hip hop, relaxed half-sung vocal, 80 BPM, dusty drums, mellow electric piano, vinyl texture.

• Cinematic pop

  English, cinematic pop, expressive alto, 90 BPM, piano, swelling strings, deep percussion, slow build to a full final chorus.

• Metal

  English, melodic metal, powerful male vocal, 160 BPM, palm-muted guitars, double-kick drums, layered chorus vocals.

YuE2 builds full song structures — intro, verse, chorus, bridge, outro — and the section tags are how you ask for them. Put each tag on its own line in square brackets and leave a blank line between sections.

[Verse]
Porch light on, the street is still
Coffee cooling on the sill

[Chorus]
Call it early, call it late
I will be here, I can wait

[Verse]
Radio hums a song we knew
Every line comes back to you

[Chorus]
Call it early, call it late
I will be here, I can wait

[Outro]
I can wait

• Keep lines short and singable, with a similar syllable count inside a section.
• Repeat the chorus text exactly when you want the hook to return.
• Use a bridge only when the song needs a change of perspective.
• The model does not write lyrics for you in this studio; an empty lyrics field is rejected.

In this studio the style field holds up to 2,000 characters and the lyrics field up to 5,000. There is no exact duration setting: the length of the track follows the structure of the lyrics. The planning mode (full, melody or off) and the ABC score are available in ComfyUI and the local runtime, not in the hosted workspace.

1. Production notes in the lyrics

   Lines such as “guitar solo here” or “fade out slowly” belong in the style brief. In the lyrics they can be sung as words.

2. A prose paragraph instead of lines

   Tags and line breaks are how the structure is communicated. Break the text into sections first.

3. Too many genres

   A long list of styles gives the model conflicting directions. Choose one and name its instruments.

4. No language

   State the singing language in the style, especially when the lyrics are short.

5. Changing everything between takes

   Revise one variable — tempo, lead instrument, vocal tone, or the chorus — so the next result answers a question.

6. Borrowed material

   Describe the sound you want rather than naming an artist, and submit only lyrics and reference audio you have the right to use.