Live Music Diffusion Models for Efficient Interactive Generation

A new research paper proposes Live Music Diffusion Models (LMDMs), a modification to audio diffusion models that enables efficient interactive streaming music generation on consumer hardware. The study identifies critical inefficiencies in block-wise outpainting diffusion pipelines, which previously made diffusion models computationally worse than discrete autoregressive (AR) counterparts. By simplifying the generative diffusion process, LMDMs recover and surpass the inference complexity of discrete Live Music models, making interactive music generation accessible without industrial-scale compute. The work leverages the open-source community's support for diffusion models while addressing their non-streaming bidirectional nature. The paper is published on arXiv under identifier 2605.22717.