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

Overview

This page hosts listening examples and sample interactions for Live Music Diffusion Models: Efficient Fine-Tuning and Post-Training of Interactive Diffusion Music Generators. Curated highlights appear first (cherry-picked text-to-music samples and live generative delay demos), followed by uncurated evaluation samples.

Cherry-picked examples

A curated set of text-to-music generations spanning diverse genres, tempos, and moods, selected to showcase the model's range.

LMDM Generative Delay Interactions

Sessions used the real-time generative delay deployment described in the paper; screen captures are hosted on YouTube.

Text-conditioned generation

Uncurated comparison samples. Each row is one text prompt. Columns compare the fine-tuned LMDM variants against the fine-tuned + ARC-forced LMDM variants in block-causal (BC) and encoder-decoder (ED) configurations. Each cell shows two samples: primed (model given an audio context prefix) and unprimed (generation from scratch).

Time-varying (prompt transitions)

Each clip is generated by LMDM (Enc–Dec, ARC-forced, block size 192/48) with a mid-piece prompt switch from A → B. The model receives the new text condition at the transition point and must smoothly continue generation.

Accompaniment generation

Each row is a different accompaniment example. Columns vary the stem lookahead: how far into the future of the conditioning stem the model can see at inference time.

+2 s lookahead

The model sees 2 seconds ahead of the current playback position in the stem it accompanies.

0 s lookahead

The model sees the stem only up to the current playback position, with no future information.

−2 s lookahead

The model sees the stem only up to 2 seconds before the current playback position; fully causal with an added buffer delay.

How to cite

If you use this work, please cite our paper.

@article{novack2026lmdm,
  title         = {Live Music Diffusion Models: Efficient Fine-Tuning and Post-Training of Interactive Diffusion Music Generators},
  author        = {Novack, Zachary and Brade, Stephen and Kim, Haven and Flores Garc{\'i}a, Hugo and Shikarpur, Nithya and Talegaonkar, Chinmay and Kim, Suwan and Chen, Valerie K. and McAuley, Julian and Berg-Kirkpatrick, Taylor and Huang, Cheng-Zhi Anna},
  journal       = {arXiv preprint arXiv:2605.22717},
  year          = {2026},
  archivePrefix = {arXiv},
  eprint        = {2605.22717},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2605.22717}
}