AIRGen

Offical repository of Arrange, Inpaint, and Refine: Steerable Long-term Music Audio Generation and Editing via Content-based Controls. You can visit our demo page here.

Quick Start

Requirements

The project requires Python 3.11. See requirements.txt.

conda create -n <environment-name> --file requirements.txt

Inference

python inference.py --output_folder=your_output_folder \
                    --model_path=your_model_weight_path \
                    --mode=fix_drums \
                    --audio_path=your_audio_path \
                    --midi_path=your_midi_path \
                    --chord_path=your_chord_path \
                    --beat_path=your_beat_path \
                    --drums_path=your_drums_path \
                    --onset=your_onset

See demo_fm folder for the input data format. Regarding mode, you can choose one from {fix_drums, edit_drums, piano, chord}. Onset should be a number indicating the starting second of the input audio. You can download soundfont here.

Model Weights

We provide with model weights (3.75M) with m=10. Download them via drums, chord, and piano.

How to cite

@misc{lin2024arrange,
title={Arrange, Inpaint, and Refine: Steerable Long-term Music Audio Generation and Editing via Content-based Controls}, 
author={Liwei Lin and Gus Xia and Yixiao Zhang and Junyan Jiang},
year={2024},
eprint={2402.09508},
archivePrefix={arXiv},
primaryClass={cs.SD}
}