Recently, on January 30, 2025, YuE transitioned to the Apache 2.0 license, marking a pivotal moment in its development. This change allows for broader use and adaptation of the model, potentially catalyzing innovation in music generation AI similar to what Stable Diffusion and LLaMA achieved in their respective domains. The developers actively encourage artists and content creators to incorporate YuE-generated outputs into their works, even for commercial purposes, with the simple requirement of crediting "YuE by HKUST/M-A-P".
YuE's architecture is designed to handle complex music generation tasks, requiring substantial computational resources. For optimal performance, the developers recommend using GPUs with at least 80GB of memory for full song generation, though it can run on smaller setups with some limitations. The model's execution time varies depending on the hardware used, with an H800 GPU capable of generating 30 seconds of audio in about 150 seconds, while an RTX 4090 GPU takes approximately 360 seconds for the same task.
Key features of YuE include:
- Full-song generation capability
- Support for both Chain of Thought (CoT) and In-Context Learning (ICL) modes
- Ability to generate music based on user-provided lyrics and genre specifications
- Optional audio prompt feature for style-based generation
- Flexible output options, including the ability to generate multiple song segments
- Compatibility with popular deep learning frameworks and tools
- Open-source nature, allowing for community contributions and improvements
- Automatic download of language model checkpoints from Hugging Face
- Support for Flash Attention 2 to optimize GPU memory usage
- Customizable inference parameters for tailored music generation
YuE represents a significant step forward in AI-driven music creation, offering a powerful and flexible tool for artists, producers, and enthusiasts to explore new realms of musical creativity.