YuE2 Prompting Guide: Style Tags, Lyrics, and Planning Modes

YuE2 is one of the two AI music models built into Song Creator Pro, and it's the one that made headlines in September 2026 by outscoring Suno on the WildSongBench benchmark, beating v5 on the very day Suno retired it, and outscoring the replacement v6 when both keep the better of two takes. In Song Creator Pro it does two things: generates complete structured songs from your lyrics (verses, choruses, bridges, full accompaniment, and vocals in English, Mandarin, or Japanese), and remixes an existing song into a new style you describe.

Using YuE2 is as simple as any other generator: write a style prompt and lyrics, hit generate, get a song. But under the hood it works differently, and knowing how pays off. Instead of going straight from your text to audio, YuE2 first writes a symbolic plan, the actual melody and chords of your song, and then performs that plan. You never have to touch the plan; it's invisible unless you go looking. But it's why YuE2's song structure is so reliable, it's why the formatting rules below matter more than they do elsewhere, and if you're a power user, you can open the score and adjust the melody, chords, or lyrics before re-rendering. This guide covers everything: style prompts, lyric formatting, planning modes, the remix workflow, and the mistakes that quietly ruin generations.

New to the model itself? Our overview of what YuE2 is covers who made it, how the score planning works, and what it can do. If you're using the ACE-Step 1.5 model instead, see our AI Music Prompting Guide. The two models read prompts differently.

The Quick Version

1. Select YuE2 as your model, then write a style prompt as comma-separated tags: genre, instruments, mood, vocal gender, vocal tone. Put the most important words first.
2. Label every lyric section in square brackets: [Verse], [Chorus], [Bridge]. Separate sections with a blank line.
3. Write repeated choruses out in full every time they appear. YuE2 doesn't expand "(repeat chorus)".
4. Keep everything that isn't sung out of the lyrics field. No titles, no production notes, no "(drums build here)".
5. Match your lyric language to your style prompt. "Japanese city pop" needs Japanese lyrics.
6. Generate 2-4 candidates and keep the best. Best-of-N is how YuE2 beats Suno v5 on benchmarks, and batch generation makes it one click.

That gets you most of the way. The rest of this guide explains why these rules exist and how to push further.

How YuE2 Reads Your Prompt

Understanding the pipeline makes every rule below obvious.

When you hit generate, YuE2 works in two stages:

1. Planning. The model reads your style prompt and lyrics and composes a symbolic score: the melody line, the chord progression, and how your lyric sections map onto the song's structure.
2. Rendering. It then performs that score as 48kHz stereo audio, with vocals singing your exact words.

Both stages run automatically from a single click of generate. You don't see or manage the plan unless you choose to open it, and most tracks never need you to.

This is why YuE2's structure is so dependable: your [Chorus] tag isn't a vague suggestion, it becomes an actual planned section of the composition. It's also why sloppy formatting hurts more here than with other models. Anything you put in the lyrics field is treated as words to be sung and planned around. A stray "(guitar solo)" in your lyrics isn't a stage direction to YuE2; it's four syllables the model will try to set to melody.

Writing Style Prompts

The style prompt defines everything about the sound. YuE2 responds best to comma-separated tags covering five ingredients:

Ingredient	Example tags
Genre	pop, city pop, jazz-funk, cyber metal, ballad, lo-fi hip hop
Instruments	acoustic guitar, Rhodes piano, pulsing bass, analog synths, brushed drums
Mood	dreamy, uplifting, melancholic, energetic, intimate
Vocal gender	female vocal, male vocal
Vocal tone	airy vocal, deep voice, bright vocal, warm lead vocal, expressive lead vocal
Tempo (optional)	88 BPM, 96 BPM, upbeat, slow, unhurried phrasing

Add the language as a tag when you're not writing in English. A plain "Japanese" or "Mandarin" tag is all it takes.

Specify BPM explicitly when tempo matters to you. A BPM tag appears in the developers' own example prompts. Treat it as guidance rather than a metronome guarantee.

Exclude an instrument with a negative tag. "No guitar" shows up in the developers' own examples, and it's the cleanest way to keep an instrument the genre usually implies out of your mix.

Order matters, but loosely: earlier words get more weight. This one is a community finding rather than an official rule, but it's held up consistently in testing. Lead with whatever you care about most. If the genre is non-negotiable, put it first. If the vocal sound defines the track, lead with that.

Three example prompts you can adapt:

jazz-funk, warm lead vocal, Rhodes piano, electric bass, tight drums, upbeat

Japanese, city pop, upbeat, danceable, groovy bass, bright female vocal, 80s polish

slow neo-soul, electric piano, round bass, restrained drums, smoky female vocal, intimate late-night mood

Notice what these have in common: every tag is concrete and audible. "Groovy bass" describes a sound. "A song about heartbreak" does not; that belongs in your lyrics, not your style prompt.

Style Prompt Mistakes

• Don't describe the story. The style field is for how the song sounds, not what it's about. Themes, narrative, and emotion-through-words all belong in the lyrics.
• Don't stack conflicting genres. "Death metal, gentle acoustic folk" forces the planner to compromise on both. If you want a genre blend, pick an established hybrid ("folk rock") or lead decisively with one genre.
• Don't forget the vocal tags. If you don't specify vocal gender and tone, you're rolling dice on the most recognizable element of the track.

Formatting Lyrics

YuE2 reads your lyrics as a sequence of labeled sections, not as one long poem. Two rules come straight from the developers:

Label every section in square brackets, on its own line. Standard labels work best: [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Interlude], [Outro].

Separate sections with a blank line. One empty line between every section.

The next four aren't in the official docs. They were worked out by early users in the weeks after release, and they hold up well in practice:

Keep sections to a handful of lines. Users consistently report each section translating to roughly 30 seconds of music. Four to six lines per section is a reliable range.

Write out every repetition in full. When the chorus comes back, paste the complete chorus text under a fresh [Chorus] label. YuE2 does not expand shorthand like "(repeat chorus)" or "x2"; it will try to sing those words.

Keep repeated lines identical in syllable count. This is the most-cited trick among early users, and it makes sense given the pipeline: if your chorus melody is planned around a 9-syllable line, changing it to 12 syllables in the second chorus forces rushed, misaligned delivery. When a line repeats, repeat it exactly, or keep the syllable count matched if you vary the words.

Start with a verse or chorus, not a lyric-filled intro. If you want an instrumental opening, use an empty [Intro] label with no words under it.

A correctly formatted example:

[Intro]

[Verse]
Streetlights paint the empty road in gold
Every promise that we made is getting old
I keep your photograph inside my coat
Reading messages you never wrote

[Chorus]
So I'm driving through the night to find you
Even if the map says turn around
Every mile is one more thing I can't undo
But your voice is still the only sound

[Verse]
Coffee going cold at 4 AM
Wondering if I'd do it all again
Every exit sign's a second chance
I keep missing them to keep this dance

[Chorus]
So I'm driving through the night to find you
Even if the map says turn around
Every mile is one more thing I can't undo
But your voice is still the only sound

[Outro]
Your voice is still the only sound

The One Rule That Matters Most

The lyrics field should contain nothing but section tags and words that get sung. This one is official; the developers state it directly in their docs. Every common lyric mistake is a violation of this rule:

• Production notes: "(drums build up here)", "(pause)", "(softly)"
• The song title written above the lyrics
• Shorthand: "(repeat chorus)", "(chorus x2)"
• Hyphenated syllable forcing like "ne-ver-more" (unreliable with YuE2; write words normally)

YuE2 will attempt to sing anything that isn't a section tag. If it shouldn't be sung, it shouldn't be in the field.

Matching Language to Style

YuE2 sings in English, Mandarin, and Japanese, and it works best when your style prompt and lyrics agree. If your style says "Mandarin ballad", write the lyrics in Mandarin. If it says "English rock", write them in English. A mismatch (an "English rock" prompt with Japanese lyrics) produces less natural phrasing and pronunciation.

If you're translating existing lyrics into another language, don't translate word-for-word. Adapt the line so the syllable count, stresses, and breathing points still fit the melody. A line that scans naturally in English often needs restructuring to sing naturally in Japanese.

For languages beyond these three, switch to the ACE-Step 1.5 model, which supports 50+ languages.

Planning Modes

YuE2 exposes a choice about how much symbolic planning happens before audio is rendered. Song Creator Pro surfaces this as the planning mode setting:

Mode	What it does	When to use it
Full (default)	Plans both melody and chords before rendering, and the resulting score is editable	Original songs. This is the right choice almost always.
Melody	Plans only the melody, leaves the accompaniment free	Cases where you want the arrangement loose around a fixed melody
Off	Skips planning entirely and renders directly	Quick experiments where structure doesn't matter

This setting applies to songs you generate from scratch; remixes configure their planning automatically, so there's nothing to set there. Leave it on full unless you have a reason not to. Full planning is what makes YuE2's structure reliable and its songs editable after the fact: you can adjust the melody, chords, or lyrics in the plan and regenerate the audio without starting over.

Instrumentals

The official docs don't cover instrumentals, but the community converged on an approach that works reliably:

1. Use empty section labels in the lyrics field, so the model still plans a structure:

[Intro]

[Verse]

[Chorus]

[Verse]

[Chorus]

[Outro]

2. Remove every vocal reference from your style prompt. No "female vocal", no "airy vocal". Describe only instruments, genre, and mood.
3. Keep planning mode on full so the melody still gets composed; it will be carried by an instrument instead of a voice.

Remixing an Existing Song

Alongside writing songs from scratch, Song Creator Pro supports YuE2's other headline capability: remixing. You give it a song and a style prompt, and YuE2 rebuilds the track in the style you described while keeping the song recognizable.

There's one wrinkle worth understanding: YuE2 needs the original song's lyrics to plan a remix. Song Creator Pro handles this by auto-transcribing the input song, and the transcribed lyrics land in the lyrics field before generation. That leaves you two things to get right:

1. Review the transcription. Auto-transcription is good but not perfect, and every rule in this guide still applies: whatever sits in the lyrics field is what gets sung. A misheard word becomes a sung misheard word. Skim the transcribed lyrics and fix any errors before you generate.
2. Write the style prompt for the destination, not the source. Describe the sound you want the remix to have, using the same tag structure as any other style prompt: genre, instruments, mood, vocal character. There's no need to describe the original song; that's what the audio input is for.

Planning mode is configured automatically for remixes, so the style prompt and the transcribed lyrics are the only inputs you control. Same rule as always: if the lyrics field is clean and the style prompt is concrete, the remix will follow.

For the full walkthrough, including the one-time cover tools download, editing transcriptions, and covers with rewritten lyrics, see How to Remix a Song with YuE2.

Generate Multiple Candidates. Always.

Here's the most important workflow habit, backed by the benchmark data: on WildSongBench, standard YuE2 (the better of two takes) scores 6.73 and best-of-8 scores 6.96. For context, Suno's since-retired v5 scored 6.87 and its current v6 scores 6.56 with the same best-of-two treatment. The model's ceiling is higher than its average, and you reach the ceiling by generating several takes and keeping the best.

In Song Creator Pro, set batch size to 2-4 and let batch generation run the candidates in one click. Listen, keep the winner, discard the rest. Since generation is local and unlimited, extra candidates cost nothing but time. This single habit does more for your results than any prompt refinement.

YuE2 or ACE-Step: Which Model for This Track?

Song Creator Pro includes both models. A quick decision guide:

You want	Use
A complete song from a description and lyrics	Either; both models are strong here
To remix an existing song into a new style	YuE2
To edit melody or chords after generating	YuE2
Vocals in Mandarin or Japanese	YuE2
To extract vocals or instruments from a song	ACE-Step 1.5
To revise one section of a song without regenerating the rest	ACE-Step 1.5
Vocals in a language beyond English, Mandarin, Japanese	ACE-Step 1.5
Fast background music and rapid iteration	ACE-Step 1.5

The prompting styles differ too: ACE-Step wants dense descriptive detail and gives you knobs like guidance scale, while YuE2 wants clean tag-based style prompts and strictly formatted lyrics, and rewards you with dependable structure. If you've been using ACE-Step, the biggest adjustments are writing choruses out in full and keeping everything unsingable out of the lyrics field.

A Complete Example to Try

Model: YuE2

Style prompt:

city pop, upbeat, danceable, groovy bass, bright female vocal, 80s polish

Lyrics: use the driving-through-the-night example from the formatting section above.

Settings: planning mode on full, batch size 2. Generate, compare the two takes, keep the better one. Then try one targeted change (swap "bright female vocal" for "warm male vocal", or rewrite the bridge) and generate again.