Specialized tools

Create a track with Google Lyria 3

Choose the Pro or Clip model, write a structured prompt, optionally include lyrics and reference images, then monitor the generated track.

Follow this guide in the studio ↗

Open Tools → Google Lyria 3. This is a separate generation workspace with its own models, prompt format, gallery, and credit rules; settings from the main AI Music Generator are not carried over.

Choose model and output

The current page offers:

  • Lyria 3 Pro for tracks up to about three minutes and a higher credit cost;
  • Lyria 3 Clip for short results up to about 30 seconds and a lower cost.

MP3 is the active output format. WAV appears in the interface but is currently marked as pending, so do not plan a workflow around it until the control becomes available.

Add a working title, then decide whether to enable With lyrics. When enabled, put the lyric text and the musical instructions in clearly separated parts of the prompt.

Write a prompt that can be interpreted

Use short labeled lines instead of one dense paragraph:

Genre: nocturnal synth-pop
Mood: restrained, then euphoric
Tempo: 112 BPM
Instruments: pulsing bass, dry drums, glassy arpeggio
Structure: sparse verse, wide chorus, short instrumental outro

Use Load default prompt to inspect the expected shape, then replace it with your own direction. The page also accepts up to 10 reference images in PNG, JPEG, or WebP. Treat images as mood and visual-context references; the written prompt should still explain the music.

Generate and inspect the result

Before choosing Generate, review the summary chips for model, lyrics state, format, and image count. The page lists recent tracks and polls unfinished work automatically.

When a result is ready, listen for adherence to form, tempo feel, instrumentation, and lyric handling. Compare it only with results made for the same brief; changing the model, lyrics, and images at once makes the comparison uninformative.

Record the model, prompt, lyric state, reference images, and output date beside every approved result so a later iteration has a reliable baseline.

Review before spending another generation

First verify that the result used the intended model and lyric setting. Make timestamped notes for the opening, section changes, strongest moment, and ending. Separate instruction failures from taste: a missing requested instrument calls for a clearer prompt, while a correctly followed direction that feels wrong calls for a different creative decision. Keep a control version with unchanged settings. Use one image or a coherent set instead of unrelated references. Name each experiment by the single variable changed. This gives collaborators a precise reason to approve or reject a result.

Sources and further reading

Product interfaces change quickly. These sources were checked on the verification date above.

  1. Current product studio — interface and operation path ↗
  2. Google DeepMind — Lyria ↗