MelodyFlow

MelodyFlow is a text-controlled high-fidelity music generation and editing model that uses continuous latent representation sequences to avoid the information loss problem of discrete representations. The model is based on a diffusion transformer architecture and trained with flow matching objectives to generate and edit diverse high-quality stereo samples with the simplicity of textual descriptions. MelodyFlow also explores a new regularized latent inversion method for text-guided editing on zero-shot testing and demonstrates its superior performance on a variety of music editing cues. The model is evaluated on both objective and subjective metrics, demonstrating comparable quality and efficiency to the evaluation baseline on standard text-to-music benchmarks, and surpassing previous state-of-the-art techniques in music editing.

MelodyFlow is aimed at music producers, composers, audio engineers, and any individual interested in music creation and editing. It is particularly suitable for users who wish to generate or edit music through simple text descriptions, as it provides an intuitive and efficient way to achieve music creation and modification without in-depth knowledge of music theory.