What is Lyria 3.5? The AI model that writes full songs

Lyria 3.5 is the newest music generation model in Google DeepMind’s Lyria family, and it writes complete songs rather than loops. Give it one text prompt, or an image, and it returns a finished track that runs a couple of minutes. You get an intro that sets the mood, verses, a chorus that comes back, a bridge, and an ending that resolves. Vocals and lyrics are included by default, in the language you prompt in. It is live in the Picsart AI Playground, alongside the rest of the Lyria family. You direct a track from the same prompt box you use for images and video.

What separates 3.5 from the clip generators is that it plans before it plays. It settles the arrangement first, deciding where the chorus lands and how the energy moves. Then it performs that plan as audio, at 44.1 kHz in stereo. Here are the controls that matter and what each one changes.

Composition, not generation

Most AI music tools generate sound. Lyria 3.5 composes a song, and the difference shows up in the middle of the track.

A short clip only has to establish a mood and hold it for thirty seconds. A two-minute song has to go somewhere. It needs contrast between a verse and a chorus, a moment that lifts, and a way back down. Lyria 3.5 maps those sections and their energy before any audio exists.

So you can describe a song the way a musician would and the model will honor it. Ask it to open on unaccompanied piano, let the first verse arrive loud, cut everything dead for a beat, then bring the chorus in at full size. You get that shape. Ask it to hold the bridge, strip the band away, and start the next chorus on voices alone, and that is what returns.

The Lyria family in Picsart

Three Lyria entries sit in the Playground model list, and they are built for different jobs. Lyria 3 Clip is the fast one, made for short clips, loops and previews from text and image prompts. Lyria 3 Pro handles extended music generation with vocals, up to 184 seconds. Lyria 3.5 is the full-length song generator, taking text and image prompts and returning a complete composition with vocals.

Use them together by drafting on Clip and committing on 3.5, so the cheap pass absorbs your iterations.

Direct the timeline second by second

The strongest control in Lyria 3.5 is the timestamped brief. Instead of describing a song in general, you write a timeline and the model builds to it.

Bracket a time range, name the section, and say what happens in it. Instruments get told when to enter, lyrics when to arrive and stop, and the chorus can be pinned to a specific second. That is what makes the model usable for anything that has to sync to picture.

Timeline brief for a two-minute track

[0:00 – 0:12] Intro: solo electric piano, soft room reverb, no drums. [0:12 – 0:40] Verse 1: brushed drums and upright bass enter. Warm female alto vocal, close and conversational, singing about leaving a city before sunrise. [0:40 – 1:05] Chorus: full band, bright horns, backing harmonies on the last line of each phrase. The lyrics turn hopeful. [1:05 – 1:30] Verse 2: same instrumentation as verse 1 plus a muted trumpet answering the vocal. [1:30 – 1:50] Bridge: everything drops except piano and voice, then the drums crash back in. [1:50 – 2:00] Outro: horns hold one chord and fade.

Copied

You do not have to be this precise. A single line like “put the drop twelve seconds in” or “bring the chorus back at the fifty second mark” works on its own. But when a track has to land a change on a cut in a video, the timeline is how you get it.

Words and voice

Singing is the default. Ask for nothing and you still get words and a voice, and you have three ways to take charge of them.

Write the lyrics, or steer them

Hand over your own by pasting them under a “Lyrics:” label and marking the parts with bracketed tags. `[Chorus]`, `[Verse 1]`, `[Bridge]`, `[Pre-chorus]`, `[Intro]` and `[Outro]` all register, and they tell the model which lines are the hook and which carry the story. To have a line answered by backing singers, wrap the repeat in parentheses, so “Hold on (hold on)” comes back as an answer vocal.

Steer the writing instead by describing what the song is about rather than supplying the text. The model needs a subject. If you do not give it one, it will infer a theme from the music description, and that may not be the one you had in mind. Say the lyrics are about a long drive home after a bad day, and that the chorus should turn toward relief. You will get a coherent set of words with a repeating hook.

You can also prompt vocal moments that are not really lyrics. Try a spoken line that recurs, a snatch of conversation before the beat arrives, or a single voice in the gap before a drop.

Describe the singer

Lyria 3.5 responds to a described voice far better than to a bare instruction like “female vocals”. The useful prompt names a gender, a timbre and a range, and adds the texture you want to hear.

These are starting points, not presets. Combine one with a delivery note and a genre and the vocal gets specific. Try a female alto with a smoky lower register, singing close to the microphone over a slow soul arrangement.

One limit is worth knowing up front. Prompts asking for a real singer’s voice are blocked by the model’s safety filters, as are requests to reproduce copyrighted lyrics. Describe the voice you want, not the person who has it.

Turn an image into a soundtrack

Lyria 3.5 takes images as input as well as text. Attach up to ten reference images in the Playground’s reference slot. The model reads the mood, palette and content of the scene, then composes music to match it.

This is the fastest route to a score for something you have already made. Stills from a shoot, frames from a video edit, a product photograph, a landscape: any of them carries enough atmosphere to work from. Add a short text instruction naming the kind of music you want, and the result fits the visuals rather than merely sitting behind them.

Image-to-music prompt

Read the palette and the weather in these frames and score them. Slow, wide pads, one cello note held underneath, no percussion at all. Around 70 BPM. Instrumental only.

Copied

Ten images is a lot of reference, so use it deliberately. A tight set that shares a look gives a clear read. A scattered set gives an average of everything, which sounds like nothing in particular.

How to generate a Lyria 3.5 track in Picsart

1. Open the Playground

Load the AI Playground and open the model picker at the top of the prompt bar.

Open AI Playground

2. Pick Lyria 3.5

Choose Lyria 3.5 from the audio models. Lyria 3 Clip and Lyria 3 Pro sit next to it if you want a quick draft or a shorter track instead.

3. Write the brief

Lead with genre, then add tempo, key, mood and instrumentation. Add a vocal description if you want singing, or say "instrumental only" if you do not. Use a timeline if the track has to hit specific moments.

4. Attach reference images

Drop up to ten images into the reference slot to have the model compose from the mood of your visuals. Skip this step for a text-only brief.

5. Generate, then listen all the way through

Judge the track on its middle and its ending, not its first ten seconds. Structure is what this model is for, so the chorus and the bridge are where you find out whether the brief worked.

6. Take it into the rest of your work

Your track lands in the same workspace as your images and video, ready to pair with an edit or feed into a longer project.

Genre, language, and instrumental tracks

Genre and instrumentation

Lead your prompt with genre and mix freely. Orchestral strings over a trap beat, surf rock played entirely on synths, a gospel choir dropped into a house track: the model handles blends. You can name an era too, like late seventies disco or early two-thousands garage, and it will chase that production style. Regional micro-genres are worth trying, though the model may only approximate the ones with a very specific local sound.

Instrumentation mostly takes care of itself, because the model picks what the genre implies. But a house track will not hand you a clarinet solo unless you ask, so name anything unexpected. Describing how instruments behave toward each other beats listing them. Try a fuzzed bass grinding underneath glassy hi-hats, or a wide synth pad well behind a guitar recorded close and dry. BPM and key both land, and adjectives like nostalgic, ethereal or dreamy shape the arrangement more than they should.

Lyrics in another language

Lyria 3.5 writes the words in whatever language you brief it in. Write the brief in French and you get French lyrics, with the vocal style and pronunciation adapted to match. Write it in English and ask for the lyrics in another language and it will follow that instruction too. That helps anyone publishing to more than one market. Music no longer has to stay in English or get dropped for a library instrumental.

Instrumental only

When vocals would get in the way, ask for an instrumental and you get one. This is the mode that matters most for content work, because a singer competes with a voiceover. Background music for a tutorial wants instruments and nothing else. So does a bed under a product demo, a loop for a game scene, or atmosphere behind a slideshow. Every other control still applies, so an instrumental can be timestamped, keyed, set to a tempo, and built around specific instruments.

Instrumental bed for a voiceover

A slow downtempo instrumental at 76 BPM in D minor. Muted electric piano, a brushed snare kept low, one sustained bass note under each chord, faint tape hiss across the whole thing. Hold the top end quiet and the dynamics even so a spoken voice sits cleanly above it. Instrumental only, no vocals.

Copied

Six steps, start to finish. Open the Picsart AI Playground and follow along.

Write the song you keep humming

Describe the track you want, down to the second if you need to, and let Lyria 3.5 handle the arrangement. Open the AI Playground and generate your first full-length song.