Zhengcong Fei Mingyuan Fan Changqian Yu Junshi Huang

Abstract

This paper explores a simple extension of diffusion-based rectified flowTransformers for text-to-music generation, termed as FluxMusic. Generally,along with design in advancedFluxhttps://github.com/black-forest-labs/flux model, we transfers itinto a latent VAE space of mel-spectrum. It involves first applying a sequenceof independent attention to the double text-music stream, followed by a stackedsingle music stream for denoised patch prediction. We employ multiplepre-trained text encoders to sufficiently capture caption semantic informationas well as inference flexibility. In between, coarse textual information, inconjunction with time step embeddings, is utilized in a modulation mechanism,while fine-grained textual details are concatenated with the music patchsequence as inputs. Through an in-depth study, we demonstrate that rectifiedflow training with an optimized architecture significantly outperformsestablished diffusion methods for the text-to-music task, as evidenced byvarious automatic metrics and human preference evaluations. Our experimentaldata, code, and model weights are made publicly available at:https://github.com/feizc/FluxMusic.