Long-form music generation with latent diffusion

Audio-based generative models for music have seen great strides recently, butso far have not managed to produce full-length music tracks with coherentmusical structure. We show that by training a generative model on long temporalcontexts it is possible to produce long-form music of up to 4m45s. Our modelconsists of a diffusion-transformer operating on a highly downsampledcontinuous latent representation (latent rate of 21.5Hz). It obtainsstate-of-the-art generations according to metrics on audio quality and promptalignment, and subjective tests reveal that it produces full-length music withcoherent structure.