MIDI-LLM: Adapting Large Language Models for Text-to-MIDI Music Generation
Sound
2025-11-07 v1 Computation and Language
Multimedia
Abstract
We present MIDI-LLM, an LLM for generating multitrack MIDI music from free-form text prompts. Our approach expands a text LLM's vocabulary to include MIDI tokens, and uses a two-stage training recipe to endow text-to-MIDI abilities. By preserving the original LLM's parameter structure, we can directly leverage the vLLM library for accelerated inference. Experiments show that MIDI-LLM achieves higher quality, better text control, and faster inference compared to the recent Text2midi model. Live demo at https://midi-llm-demo.vercel.app.
Cite
@article{arxiv.2511.03942,
title = {MIDI-LLM: Adapting Large Language Models for Text-to-MIDI Music Generation},
author = {Shih-Lun Wu and Yoon Kim and Cheng-Zhi Anna Huang},
journal= {arXiv preprint arXiv:2511.03942},
year = {2025}
}
Comments
To appear at NeurIPS 2025 Workshop on AI for Music