Implementation:Facebookresearch Audiocraft MAGNeT get pretrained

Knowledge Sources	Facebookresearch_Audiocraft MAGNeT
Domains	Audio_Generation, Masked_Generation
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for loading pretrained MAGNeT non-autoregressive text-to-music and text-to-sound generation models.

Description

MAGNeT provides a high-level user-facing API for non-autoregressive audio generation. Unlike autoregressive models, the generation duration is fixed by the model configuration and not user-configurable at inference time. The set_generation_params method configures MAGNeT-specific parameters including annealed CFG, per-codebook decoding steps, and span arrangement.

Usage

Import this class when you want to generate music or sounds using the faster non-autoregressive MAGNeT approach.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/models/magnet.py
• Lines: 1-88

Signature

class MAGNeT(BaseGenModel):
    @staticmethod
    def get_pretrained(name: str = 'facebook/magnet-small-10secs', device=None):
        """Load a pretrained MAGNeT model."""

    def set_generation_params(self, use_sampling=True, top_k=0, top_p=0.9,
                              temperature=3.0, max_cfg_coef=10.0, min_cfg_coef=1.0,
                              decoding_steps=[20,10,10,10], span_arrangement='nonoverlap'):
        """Configure MAGNeT generation parameters."""

Import

from audiocraft.models import MAGNeT

I/O Contract

Inputs

Name	Type	Required	Description
descriptions	list[str]	Yes	Text descriptions for generation
name	str	No	Pretrained model name

Outputs

Name	Type	Description
wav	torch.Tensor	Generated audio [B, C, T]

Usage Examples

from audiocraft.models import MAGNeT

model = MAGNeT.get_pretrained('facebook/magnet-small-10secs')
model.set_generation_params(temperature=3.0, top_p=0.9)
wav = model.generate(['upbeat electronic music with heavy bass'])

Related Pages

• Principle:Facebookresearch_Audiocraft_Masked_Parallel_Token_Generation