MAGNeT

Meta’s non-autoregressive masked transformer for text-to-music and text-to-sound — much faster than MusicGen.

Strengths

• Non-autoregressive — generates in a fixed number of steps rather than token by token, so it is far faster than MusicGen
• Single-stage: no semantic conditioning, model cascading or audio prompting required
• Both music and sound-effect checkpoints, at 10 and 30 seconds
• Same AudioCraft API as MusicGen and AudioGen

Limitations

• Non-commercial weights (CC-BY-NC-4.0)
• Fixed output lengths per checkpoint — pick 10 s or 30 s at load time
• Quality generally trails MusicGen at the same size; speed is the trade
• Less widely used, so fewer worked examples exist

Best for

Iterating quickly on text-to-audio ideas, or any workflow where MusicGen’s sequential generation is the bottleneck.

Minimal example

from audiocraft.models import MAGNeT
from audiocraft.data.audio import audio_write

model = MAGNeT.get_pretrained('facebook/magnet-small-10secs')
wav = model.generate(['80s electronic track with melodic synthesizers'])

audio_write('magnet_out', wav[0].cpu(), model.sample_rate, strategy='loudness')

Checkpoints

Related tools

See all MAGNeT alternatives & how they compare →

See the whole ML & generative category →

Head to head

← Back to all tools