Gradio

Audio Duration (-1 for automatic)

Length of the generated song, in seconds.

-1 240

Tags

The style to generate: genre, instrumentation, mood, tempo. This is what steers the sound most.

Lyrics

The words to sing. Leave it empty for an instrumental.

Infer Steps

More steps give a cleaner result and take longer.

1 200

Guidance Scale

How closely the result follows the tags and lyrics. Higher values track them more literally and tend to sound less natural.

0 30

Granularity Scale

Controls how much fine detail the model commits to. Raise it for a busier, more detailed arrangement.

-100 100

manual seeds (default None)

Fixes the result, so the same seed and settings give the same song. Leave it empty for a new one each run.

Guidance Interval

The portion of generation over which guidance is applied. Narrower intervals leave the model freer at the start and end.

0 1

Guidance Interval Decay

How quickly guidance falls off across that interval. Higher values let the ending wander further from the prompt.

0 1

Min Guidance Scale

The floor guidance decays to, so the prompt is never ignored entirely.

0 200

use ERG for tag

Applies the model's tag-conditioning refinement. Leave off unless the style is being ignored.

use ERG for lyric

Applies the same refinement to the lyrics, which can improve how clearly words are sung.

use ERG for diffusion

Applies the refinement during diffusion. On by default, and turning it off is faster but rougher.

OSS Steps

Advanced: a comma-separated list of step indices to resample. Leave it empty unless you are reproducing a specific result.

Guidance Scale Text

Weighs the text prompt separately from the lyrics. Raise it to favour the described style.

0 10

Guidance Scale Lyric

Weighs the lyrics separately from the style. Raise it to favour clearly sung words.

0 10

Enable Audio2Audio

Uses the reference track below as a starting point instead of generating from nothing.

Refer audio strength

How much of the reference track survives. Low values keep more of it, high values depart further.

0 1

Reference Audio (for Audio2Audio)

Lora Name or Path

An optional fine-tune to load on top of the base model.

Text2Music Generated Audio 1