MAGNeT
verifiedFreehuji.ac.il
Meta research model (MAGNeT) for fast text-to-music/audio via a single non-autoregressive transformer; a demo and paper, not a product.
What it does
MAGNeT is a research text-to-music and text-to-audio model from Meta's FAIR team and collaborators. It uses a single-stage, non-autoregressive transformer over audio tokens to generate high-quality audio much faster than autoregressive baselines, with published code and a paper.
Core features
Single non-autoregressive transformer
Text-to-music and text-to-audio generation
About 7x faster than autoregressive baselines
Novel rescoring for higher quality
Hybrid autoregressive/non-autoregressive variant
Open code and paper
Best for
→Generate music from text prompts
→Produce audio samples for research
→Study non-autoregressive audio generation