How MusicGen takes direction
MusicGen takes a free-text description — genre, instruments, mood, tempo — and generates instrumental audio. There is no lyrics field and no vocal mode; the released models are trained on instrumental music. Your control comes from the description plus sampling settings: temperature, top_k, top_p and classifier-free guidance. Model size trades speed for fidelity, and the melody variant conditions on a monophonic melody you supply yourself.
Layer by layer
- Front-load the description. Genre and instruments first; the model weights the opening words most heavily.
- Use numbers where you have them. A BPM and a key land better than "mid-tempo" and "melancholy".
- Plan in 30-second windows. Write the continuation as its own description instead of expecting a three-minute arc from one pass.
- Keep temperature near 1.0. Around 0.9–1.1 is the usable band; above 1.3 it drifts, far below 0.7 it loops.
- Set cfg_coef around 3. Higher values obey the description more literally and sound less natural.
- Fix the seed when comparing. Two descriptions plus two random seeds tells you nothing about either description.
- Melody conditioning is for your own material. Hum a line, render it, condition on that — not on someone else's recording.
MusicGen setups you can paste
One description, one pass, then continue from the tail if the track needs to be longer.
Lo-fi study loop · 82 BPM
Cinematic build · 110 BPM
Ambient pad · 64 BPM
More ready-made prompts across moods and engines: AI music prompt examples.
What to leave out
- Expecting vocals. The released MusicGen models are instrumental. Use a song model if the track needs a voice.
- A paragraph as a description. One line, front-loaded, is what the model conditions on best.
- Asking one pass for three minutes. Thirty seconds is the unit of work; plan the structure across passes.
- Melody conditioning from a commercial recording. Supply a melody you wrote, or stay with text.
- Ignoring the licence. The code is permissively licensed; the published weights carry a non-commercial condition.
The rule underneath all of them: describe the music, never the record. Reference an artist or a track and you inherit someone else's rights problem along with a vaguer brief.
Build a MusicGen brief from a reference track
The hard part is the honest numbers: tempo, key, brightness, dynamics, and the instruments actually carrying the groove. Music DNA Studio measures them from any MP3, WAV, FLAC or MP4 — locally in your browser, nothing uploaded — then outputs a MusicGen-shaped prompt alongside versions for the other fourteen engines. You make the creative calls; the spec underneath is measured, not guessed.
Because the description is one line, every word has to earn its place — which is exactly what measured numbers are good for.
Guides for the other engines
Eleven Music · Suno · Udio · Mureka · MiniMax Music · YuE2 · Soundverse · Treblo · Stable Audio · Soundraw · Mubert · Beatoven · AIVA · Loudly · MusicGen
Build The MusicGen Brief
Drop in a reference, pick MusicGen, copy the description and sampling settings. Free, no signup, nothing leaves your device.
ANALYSE YOUR AUDIO — FREE →FAQ
Can MusicGen generate vocals? ▾
No. The released AudioCraft music models generate instrumental audio and have no lyrics or vocal field. For a sung track you need a song model; the measured brief transfers, the engine does not.
How long can one generation be? ▾
Thirty seconds per pass. Longer pieces are built by continuing from the tail, and the cleanest joins come from overlapping a second or two and crossfading.
Do I need a GPU? ▾
For local runs, yes — the larger models are heavy. A hosted Space is the quick way to test a description before committing to hardware.
Which model size should I use? ▾
Small for fast iteration on wording, medium for a balance, large for the final render. Decide the description on small, then spend the time on large.
Can I use MusicGen output commercially? ▾
Check the weights licence before you plan a release: the code and the published model weights are licensed differently, and the weights carry a non-commercial condition.