Reproduce “Mamba-Diffusion Model with Learnable Wavelet for Controllable Symbolic Music Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jincheng Zhang, György Fazekas, Charalampos Saitis
Proffusion-WM is a diffusion model for controllable symbolic music generation that combines Mamba selective state space models with learnable wavelet transforms. This approach generates music in pianoroll format, demonstrating improved performance over baselines in objective metrics like Overlapping Area and Chord F1 Score, and receiving higher subjective ratings in listener preference and harmonic quality.
Key insights
- Integrating learnable wavelet transforms within a U-Net diffusion model enables multi-resolution analysis, effectively capturing both global structures and fine-grained details in musical data.
- Combining Mamba's linear-scaling selective state space modeling with Transformer attention in hybrid blocks allows for efficient capture of long-range dependencies in music while maintaining local pattern recognition.
- Chord conditioning via a VAE encoder and classifier-free guidance provides effective harmonic control over generated music, balancing adherence to conditions with creative output.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.