Lyria 3.5: Full Songs From a Single Prompt

Thirty seconds of AI music is easy now. A loop, a sting, a mood under a product shot. You type a vibe and something pleasant comes back.

Now ask for a song. A real one, with a verse that sets something up, a chorus that pays it off, and a bridge that goes somewhere new before the last chorus lands. That's where most generated music quietly falls apart. The second verse forgets the first. The chorus never quite arrives. The track just keeps going, the same eight bars wearing different hats, until it stops.

Lyria 3.5 is Google's answer to that problem, and it's live on each::labs. The interesting part isn't that it sounds good, although Google calls it its best-sounding music generation model. It's that it's built to hold a song together from the first bar to the last.

What Lyria 3.5 Actually Is

Lyria 3.5 is Google's flagship music model. On its Gemini API model page, Google describes it as optimized for complex structural coherence: multiple verses, choruses and bridges inside one generation. It outputs 44.1 kHz stereo audio as an MP3, plus the lyrics as text, and it takes text and images as input. The model is marked stable, not preview.

Google's September 4 announcement adds the texture: more expressive vocals and richer musical arrangements than before. The same model now sits behind music features in the Gemini app, Flow Music, Google Vids and AI Studio. The use cases Google names are refreshingly unglamorous: backing tracks for video, brand jingles, ringtones. That's the right list. Most music generation isn't for the charts. It's for the forty other places a project needs sound.

On each::labs the request is deliberately simple. You send a prompt of up to 20,000 characters and, optionally, up to 10 image_urls. That's the whole schema. No genre dropdown, no duration slider, no tempo field. Everything lives in the prompt, which sounds limiting until you see how much a prompt can carry.

A Song Needs a Bridge, Not Just a Loop

Here's the position worth taking before your first generation: length isn't the hard part of music generation. Structure is.

Any model can extend audio. Making four minutes feel like a journey rather than a long loop requires the model to know where it is in the song and what that section is for. Google's music generation guide draws the line clearly between its two current models. Lyria 3 Clip Preview produces fixed 30-second clips, made for loops and previews. Lyria 3.5 produces full-length songs, typically a couple of minutes, and you control the length through the prompt itself.

That split is useful. A 30-second clip is a sketch. A full song is an argument. You want both, at different moments, and the trick is not to confuse them.

Write the Map: Section Tags and Timestamps

Because the schema has no structure fields, the prompt is the score. Google's guide supports two ways to lay out a song, and you'll get far better results using them than describing the track in one breathless paragraph.

The first is section tags. Write [Verse], [Chorus] and [Bridge] into the prompt and the model treats them as the skeleton. This is the fastest way to get a track that actually has a chorus instead of one that merely gets louder.

Upbeat indie pop, 118 BPM, key of A major, bright jangly electric guitars, warm bass, crisp live drums, female lead vocal with light harmonies. [Verse] Quiet, close vocal over muted guitar. [Chorus] Full band, wide stereo guitars, big gang vocals. [Verse] Add a tambourine and a counter melody. [Chorus] [Bridge] Half-time drums, just bass and voice, building tension. [Chorus] Final chorus with an extra harmony line, then a clean ending.

The second is timestamps, for when the timing matters more than the labels. If the music has to hit a cut in a video at 0:45, tell it so.

Cinematic orchestral score, slow build, D minor. [0:00 - 0:15] Intro: low strings and a distant piano motif. [0:15 - 0:45] Build: add French horns and soft timpani, rising tension. [0:45 - 1:15] Climax: full orchestra, choir pads, the piano motif returns in the brass. [1:15 - 1:40] Outro: strings fade back to solo piano, ending on an unresolved chord.

Google's advice is to be specific about instruments, BPM, key, mood and structure, and that's where a 20,000-character prompt stops sounding excessive. You won't need all of it. You'll need more than a sentence.

Bring Your Own Lyrics, or Ask for None

Lyria 3.5 returns lyrics as text alongside the audio, so if you don't supply words, it writes them. That's fine for a demo and rarely right for a brand jingle. If the words matter, write them yourself.

The guide's key tip is to keep your lyrics clearly separated from your instructions. A model reading "make the chorus soar" can't tell whether that's a direction or a line you want sung, unless you make the boundary obvious.

Style: warm acoustic folk, fingerpicked guitar, upright bass, brushed snare, male vocal, intimate and unhurried, around 90 BPM. Lyrics: [Verse] The kettle's on, the window's grey We've got nowhere to be today [Chorus] Slow morning, stay a while Slow morning, keep that smile

Need no vocals at all? Say it plainly: "Instrumental only, no vocals." That one line matters most for background music under voiceover, where a singing voice fighting a speaking voice ruins both.

Language follows you, too. Lyrics are generated in the language of your prompt, so a prompt written in Spanish gets you a Spanish song. For localized campaigns that's a quiet superpower: write the brief in the market's language and the song comes back already speaking it.

One practical tip saves wasted runs. Prompts that ask for a specific artist's voice or for copyrighted lyrics are blocked, so describe the sound instead of naming the singer. "Smoky, breathy alto with a slow vibrato" gets you further than any name would, and it's a better prompt anyway, because it tells the model what you actually hear.

Show It a Picture

This is the input that changes how you brief music. Lyria 3.5 accepts up to 10 images alongside the prompt, and the each::labs schema exposes all ten through image_urls.

Words are good at genre and bad at atmosphere. You can write "melancholy" a dozen ways and still not land the specific grey-blue loneliness of a rainy harbor at dusk. Hand the model the photo instead. Google's own example prompt is exactly this move:

An atmospheric ambient track inspired by the mood and colors in this image. Slow evolving synth pads, distant piano, no drums, no vocals.

For video work this is the obvious bridge. You already have frames: a storyboard, a keyframe, a still from the edit. Feed a few of them in and the soundtrack starts from the same visual world as the picture, rather than from your best attempt to describe it in words.

Draft on Clip, Finish on Lyria 3.5

The workflow Google recommends is the one you'd arrive at on your own after a week of generating: iterate on Clip first, then go full.

A 30-second Lyria 3 Clip generation is the fastest way to settle the early questions. Is this the right genre? Is 118 BPM too fast? Does the vocal tone fit the brand? Run the same style description a handful of times, keep the one that sounds right, and only then write the full structure and send it to Lyria 3.5. Both models take the same prompt-plus-images shape on each::labs, so moving between them is a model-id change, not a rewrite.

Two habits make this loop work. First, generation is single-turn: you can't ask for "same song, but slower bridge" as an edit. You change the prompt and generate again. Second, results vary between calls, so treat each prompt as a small batch rather than one shot. Run it two or three times, pick the strongest take, and keep the prompt that produced it as your template.

If you've already built prompts around Lyria 3 Pro Preview, it's live on each::labs too, with the same input shape. That makes a side-by-side easy: same prompt, two models, your ears decide.

Give Your Video a Soundtrack

Music rarely ships alone. It ships under something. The most useful thing about having Lyria 3.5 in the same catalog as the video models is that the whole piece can come from one API.

Here's a pattern that works. Generate your shots with a video model such as Seedance 2.5 Image to Video, or browse the full image-to-video lineup. Pull a couple of keyframes from the result and send them to Lyria 3.5 as image inputs. Then write the timing of your edit into the prompt with timestamps, so the build lands where the hero shot lands.

Energetic electronic track for a 60-second product video, inspired by the colors and mood in these images. 124 BPM, punchy kick, bright plucked synths, instrumental only, no vocals. [0:00 - 0:10] Intro: filtered synths, rising. [0:10 - 0:40] Main groove: full drums and bass. [0:40 - 0:50] Breakdown: drums drop out, single synth line. [0:50 - 1:00] Final hit: full mix, clean ending on the downbeat.

Want it wired together instead of stitched by hand? That's what each::labs AI Flows are for: chain the image, video and music steps once, then rerun the whole thing on a new brief.

Every Lyria output carries a SynthID audio watermark, which is worth knowing before you plan distribution. If you're shipping generated music inside a product, the licensing side deserves its own read, and our AI music commercial use checklist covers it.

Where Lyria 3.5 Fits in Your Work

Video editors and social teams get the obvious win: a backing track timed to the cut, generated against the actual frames, with no vocals stepping on the voiceover. Brand teams get jingles and sonic logos they can iterate on in an afternoon, in as many languages as they have markets. App builders get a music feature that returns a finished song and its lyrics in one response, which is everything a player UI needs.

And the small stuff counts. Ringtones, birthday songs, a podcast intro, the hold music nobody wants to license. Google built templates for background music and birthday tracks into the Gemini app for a reason. That's where most music demand actually lives.

If you're comparing music models for a long-form job, it's worth reading our take on Mureka V9.5 as well. Both models are chasing the same goal, songs that hold together, and hearing them on the same brief is the fastest way to pick.

Using Lyria 3.5 on each::labs

One endpoint, one prompt field, optional images. A call looks like this:

curl -X POST \
  -H "Authorization: Bearer $EACHLABS_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "lyria-3-5",
    "version": "0.0.1",
    "input": {
      "prompt": "Warm lo-fi hip hop, 85 BPM, dusty vinyl drums, Rhodes chords, instrumental only, no vocals. [Verse] Sparse drums and keys. [Chorus] Add bass and a soft melody. [Bridge] Drums drop, keys alone. [Chorus] Full groove, fade out.",
      "image_urls": ["https://example.com/keyframe-1.jpg"]
    }
  }' \
  https://api.eachlabs.ai/v1/prediction/

Swap the model to lyria-3-clip-preview for a 30-second draft, or lyria-3-pro-preview for comparison. Same auth, same body, same response handling. You get a prediction ID back, which you poll or receive by webhook, so build the wait into your flow from day one. A synchronous assumption breaks the first time it meets real traffic.

The rest of Google's lineup sits on the Gemini family page, next to the image and video models you'll probably pair it with.

Frequently Asked Questions

How long are Lyria 3.5 songs?

Think "a couple of minutes", which is how Google's guide puts it, and steer from there. There's no duration field on each::labs; you set length in the prompt, ideally with timestamps like "[1:30 - 2:00] Outro" so the model knows where the song ends. If you only need 30 seconds, Lyria 3 Clip Preview is the better tool.

Can Lyria 3.5 write lyrics, or do I need my own?

Both work. Leave lyrics out and the model writes them, returning the text alongside the MP3 in the language of your prompt. Write your own when the words carry the message, and keep them clearly separated from your style instructions so nothing you meant as direction ends up sung.

What's the difference between Lyria 3.5 and Lyria 3 Clip?

Shape and purpose. Clip always returns a fixed 30-second piece, made for loops, previews and fast iteration. Lyria 3.5 builds full songs with verses, choruses and bridges. The smart move is to use them together: find the sound on Clip, then commit the structure on 3.5.

Can I use images to guide the music?

Yes, up to 10 of them per request through the image_urls field. It works best for mood and color, the atmosphere words struggle with. Pair a few keyframes from your video with a short style line and the track starts from the same world as the picture.