Live Music Diffusion Models: Efficient Fine-Tuning and Post-Training of Interactive Diffusion Music Generators

Researchers introduced Live Music Diffusion Models (LMDMs) to make diffusion-based music generation interactive on consumer hardware.

The system uses block-wise KV caching to speed inference and ARC-Forcing to enable stable post-training alignment.

It was shown in text-conditioned generation, sketch-based synthesis, jamming, and a live artist-AI collaboration running locally on a gaming laptop.

The work points to a practical path for low-latency, high-quality music tools beyond offline, compute-heavy setups.