Open Weight Music Generation Models Trained on Licensed Data (stability.ai)

Stable Audio 3.0 has been launched as a new model family for music generation, offering open-weight models trained on fully licensed data. This release is significant for the AI/ML community as it fosters creativity in the audio space similar to what Stable Diffusion achieved for images. Users retain ownership of their outputs, with options to distribute and commercialize them under a supportive licensing framework, making it ideal for both individual artists and organizations. Among the innovative features of Stable Audio 3.0 are variable-length generation capabilities that allow for tracks extending up to six minutes, and the ability to compose full music on portable devices. The architecture incorporates a novel semantic-acoustic autoencoder, enabling greater flexibility and musicality. Notably, users can customize models using LoRa training, a method gaining traction in audio generation, and make on-the-fly edits to their compositions. This release represents a commitment to developing responsibly trained generative AI models that prioritize artist collaboration and community-driven innovation, setting the stage for further advancements in music technology.