| name | music-generation |
| description | Music generation. Produces a music clip from a text prompt using Lyria. Use for background scores, mood pieces, jingles, underscore tracks, or any plan node requiring generated audio. |
| metadata | {"type":"primitive"} |
When to use
Use t2m when:
- The user requests a music track, score, jingle, or background audio.
- A plan needs an audio layer to accompany video or image content.
- The promptIntent references a mood, genre, or musical style.
Prompt structure
Every music prompt must specify four core dimensions. Omitting any dimension lets the model guess, producing generic or mismatched output.
[GENRE & STYLE] [MOOD & EMOTION] [INSTRUMENTATION] [TEMPO & ENERGY]
[SONG STRUCTURE] [PRODUCTION QUALITY]
[VOCALS or "instrumental"]
[NEGATIVE PROMPT (optional)]
Core dimensions
[GENRE & STYLE]
Name the genre explicitly. Compound genres are valid and often preferred.
- Good: "cinematic orchestral," "lo-fi hip-hop," "upbeat Latin pop," "dark ambient electronic," "classic jazz trio," "progressive rock," "blues-influenced R&B"
- Bad: "nice music," "background music," "something cool"
Sub-genre precision improves output — "new-wave synth-pop" is better than "electronic."
[MOOD & EMOTION]
Use concrete emotional descriptors rather than vague adjectives.
- Good: "melancholic and introspective," "triumphant and soaring," "tense and suspenseful," "playful and whimsical," "nostalgic and warm"
- Bad: "good vibes," "emotional," "powerful"
Pair moods when appropriate: "bittersweet yet hopeful."
[INSTRUMENTATION]
Name specific instruments. Generic terms ("guitar," "strings") are weaker than specific ones.
- Good: "fingerpicked acoustic guitar, upright bass, brushed snare, Fender Rhodes," "sweeping string section with solo violin, French horns, and orchestral timpani," "808 bass, trap hi-hats, atmospheric pad synths, and sampled piano"
- Bad: "guitars and drums," "orchestra"
Always specify the lead instrument — it anchors the texture: "lead: solo cello."
[TEMPO & ENERGY]
Provide BPM when precision matters; use descriptive terms otherwise.
- Numeric: "120 BPM," "72 BPM ballad tempo," "180 BPM drum and bass"
- Descriptive: "slow and meditative," "mid-tempo groove," "driving and relentless," "rubato, free-flowing tempo"
Always pair tempo with an energy word: "slow (40 BPM), sparse and sparse" or "fast (140 BPM), energetic and propulsive."
[SONG STRUCTURE]
Optional but powerful for controlling dynamics and arc. Use only when the user has a narrative intent.
- "Builds from sparse intro → full chorus → stripped-down outro."
- "Repeating two-bar loop, no variation — suitable for seamless looping."
- "Verse-chorus-verse structure with a key change on the final chorus."
- "Single sustained ambient texture, no structural variation."
For clips (30s default model), one dynamic arc is the maximum — do not plan full song structures.
[PRODUCTION QUALITY]
Sets the sonic aesthetic.
- "studio-quality, polished mix, wide stereo field"
- "lo-fi, warm tape saturation, slight vinyl crackle"
- "raw live recording, room ambience, natural reverb"
- "cinematic, large-room reverb, IMAX-scale dynamics"
- "intimate, close-mic'd, dry acoustic"
[VOCALS]
Always declare explicitly.
- No vocals: Add
"instrumental" as a standalone keyword — always include this when vocals are not wanted.
- With vocals: Describe style, not lyrics: "wordless female vocal harmonies," "deep baritone spoken word," "layered gospel choir," "distorted punk vocals."
- Lyrics: Wrap in quotes:
vocals singing: "Rise up and take the stage."
Negative prompts
Use negative_prompt to exclude specific elements. Be precise — model-level exclusions, not abstract moods.
| Want to avoid | Negative prompt value |
|---|
| All vocals | "vocals, singing, choir, spoken word" |
| Percussion | "drums, kick, snare, hi-hat, percussion" |
| Distortion | "distorted guitar, fuzz, overdrive, noise" |
| Specific instruments | "violin, piano, brass" |
| Bass frequencies | "bass guitar, sub-bass, 808" |
| Genre bleed | "jazz, hip-hop, electronic" |
Do not use mood words in negative prompts — "sad", "boring", "cheap" have no effect.
Timestamp prompting
For dynamic genre or mood shifts within a clip, use [mm:ss] markers in the prompt. Only available for longer clips (lyria-3-pro-preview).
[0:00] Sparse piano intro, slow and contemplative, 60 BPM.
[0:15] Full orchestral swell enters — strings, brass, timpani.
[0:45] Returns to solo piano, fading out.
Rules:
- Timestamps must be in
[mm:ss] format.
- Minimum 15 seconds between transitions.
- Maximum 3–4 transition points per clip to avoid incoherence.
- Each section must still specify genre, mood, and instrumentation.
Model selection
| Model | Duration | Best for |
|---|
lyria-3-clip-preview | ~30s | Default. Loops, stingers, short background pieces. |
lyria-3-pro-preview | up to ~3min | Full tracks, complex structures, timestamp prompting. |
Always use lyria-3-clip-preview unless the user explicitly needs a longer track or timestamp-based structure.
Full prompt examples
Background underscore for a product video:
Upbeat corporate pop. Optimistic and forward-moving. Instrumentation: electric piano, clean electric guitar, light synth pads, punchy kick and snare. 118 BPM. Builds from sparse verse to full chorus at 0:15. Studio quality, polished mix. Instrumental.
Suspense cue for a horror sequence:
Dark cinematic orchestral. Tense, dread-building. Lead: low string ostinato with dissonant brass stabs. Sparse, irregular percussion. 60 BPM rubato. No melody — sustained atmospheric texture. Large-room reverb. Instrumental.
Negative prompt: piano, acoustic guitar, major key, bright tones.
Lo-fi study track:
Lo-fi hip-hop. Melancholic and cozy. Instrumentation: Fender Rhodes, vinyl-sampled double bass, brushed jazz drums, atmospheric pad. 85 BPM. Loopable, no structural variation. Warm tape saturation, vinyl crackle. Instrumental.
Cinematic epic with vocals:
Cinematic epic orchestral. Triumphant and emotionally charged. Full orchestra: soaring strings, French horns, choir, orchestral percussion. 120 BPM, driving. Builds through verse to massive final chorus. Studio quality, IMAX-scale dynamics. Wordless female and male choir.
Common failures
- Omitting "instrumental" when no vocals are wanted — the model often adds them by default.
- Using a single-word genre ("rock") without mood or instrumentation — produces a generic, unfocused result.
- Specifying lyrics without using the
vocals singing: format — lyrics may appear as narration instead.
- Using mood words in
negative_prompt — only instrument and genre terms have effect.
- Requesting complex song structure on a 30s clip — use
lyria-3-pro-preview and timestamp prompting.
- Stacking more than 4 timestamp transitions — causes abrupt, incoherent transitions.