music-gen

On the rail as music-gen.

When to use it

Making a music track from a description of its feel: a bed under a video, a loop to work to, or a full song with a vocal.

When to reach for something else

It writes from a description alone and takes no reference audio, so it cannot match an existing melody or extend a track you already have. For adding sound to footage you already shot, a different tool takes that clip directly.

How to use it

1. Write what you want in Prompt.
2. Pick a model in the pill at the top of the card, or keep Lyria 3.5.
3. Press Run.

On the card

Control	What it does	Choices
Prompt	What you want, in your own words. The card sends it with your inputs.	Your words
Run	Starts the run. The number on it is what this run costs in credits, worked out before you press.
×1	Runs the card several times in a row with the same settings.	1 to 99

Models

The pill at the top of the card picks the model. Lyria 3.5 is the one a new card starts on.

Model	Good for
Lyria 3.5	A full music track from a description alone, priced the same whatever length it runs, so a three minute song costs no more than a one minute bed. It writes real lyrics when the prompt asks for a vocal.
ElevenLabs Music v2.5	The Music tool’s default and the newer ElevenLabs weights: a full song or bed with a real arrangement, vocals unless Instrumental is on, more lifelike instruments and stronger vocal-led tracks, one to five whole minutes.
ElevenLabs Music v2	The earlier of the two ElevenLabs weights in the Music tool, same price and same minute-aligned lengths as v2.5: a full arrangement with vocals unless Instrumental is on, for a second, different take on a brief.
Sonilo v1.1	A licensed, commercial-safe music bed from a text prompt, 10 to 180 seconds and billed per second, so a short cue or an odd length costs exactly what it plays, returned as m4a.

Lyria 3.5

Not the one for: No duration control and no instrumental switch, so both live in the prompt and neither is exact: asking for thirty seconds returns about a minute, its rough floor. For an exact length, the other models take a duration directly.

This model has no settings behind the settings button.

ElevenLabs Music v2.5

Not the one for: Billed per whole minute rounded up, so a ten second sting costs a full minute; not for short cues. Text in only, no reference audio, and the Instrumental toggle is the only way to guarantee no singing.

This model has no settings behind the settings button.

ElevenLabs Music v2

Not the one for: Same whole-minute billing rounded up as v2.5, so short cues waste most of a minute, and no reference audio input. When the more lifelike instruments and stronger vocal-led tracks are wanted, those are the v2.5 weights, not these.

This model has no settings behind the settings button.

Sonilo v1.1

Not the one for: No vocal or instrumental control, only a prompt and a length, and no reference audio; when singing has to be on or off, the tool’s ElevenLabs models are the ones with the Instrumental toggle. The file is m4a, not mp3.

This model has no settings behind the settings button.