ComposerFlow: Step-by-Step Compositional Song Generation

ICLR 2026 Conference Submission18303 Authors

Keywords: Song generation, Text-to-Music, Music generation, Green AI

TL;DR: A song generation pipeline that leverages the knowledge of multiple models to achieve low-cost, fast training.

Abstract: Open-source end-to-end song generators have emerged in the wake of commercial systems like Suno, offering one-shot music with lyrics but at the cost of large datasets and substantial compute. We propose ComposerFlow, a hierarchical pipeline that composes songs by chaining four specialized components—lyrics-to-melody, melody harmonization, singing-voice synthesis (SVS), and vocal-to-backing generation. This modular design is resource-efficient (single RTX 3090; ~6k hours of data) and editable: users can revise intermediate results (e.g., melody, chords, vocal) and deterministically regenerate all downstream audio. Leveraging SVS further mitigates common end-to-end issues such as unstable vocal timbre and phoneme errors. We evaluate our pipeline against representative end-to-end baselines and find comparable perceptual quality despite significantly lower training demands. We release audio demos at https://composerflow.github.io/web/. Our results suggest that modular, controllable pipelines are a practical alternative to monolithic song models, enabling rapid iteration and reliable production on accessible hardware.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 18303