Improving Text-to-Music Generation with Human Preference Rewards

The article presents an entry to the Academic Text-to-Music (ATTM) Grand Challenge, introducing a system that integrates a learned human-preference reward from TuneJury into a 120M-parameter FluxAudio-S model. Key innovations include a training-time reward conditioning method, a variety of score-conditioning architectures, and a preference-tuning pass for improved audio-text alignment. This approach enhances text-to-music generation by leveraging human preferences, which could lead to more refined outputs in practical applications of AI music generation.