ivcylc/OpenMusic

Implements a quality-aware masked diffusion transformer (QA-MDT) that combines diffusion models with discrete token prediction, injecting quality metrics during generation to balance fidelity with musicality. Built on PyTorch with modular architecture supporting multiple backbone options (MDT, DiT, U-Net), integrates CLAP for audio understanding, FLAN-T5 for text encoding, and extends to infinite-length generation through zero-shot inference. Pre-trained checkpoints available on Hugging Face with Gradio and Diffusers implementations for straightforward deployment.

634 stars. No commits in the last 6 months.