FlowSonic: Stable Zero-Shot Music Editing via High-Order Trajectory Integration

How can we perform high-fidelity, zero-shot editing of real-world music recordings using diffusion transformers while maintaining structural integrity and semantic alignment? Existing methods often struggle with the trade-off between applying desired modifications and preserving the original musical characteristics, particularly when dealing with complex, polyphonic audio.

FlowSonic treats music editing as a two-stage process: deterministic inversion followed by conditional generation. First, the input audio is mapped into the latent space of a pretrained rectified-flow transformer. To ensure the original structure is preserved, the model reuses cross-attention representations extracted during this inversion phase.

To address the numerical instability common in inversion-based editing, the authors introduce a high-order numerical integration framework using the third-order Adams–Bashforth (AB3) solver. A key innovation is the Dynamic History Caching (DHC) mechanism, which captures velocity evaluations during the inversion stage to initialize the generation solver. This eliminates the need for lower-order warm-start initialization, which typically introduces errors that degrade the final output.

FlowSonic demonstrates that high-order numerical integration significantly improves latent trajectory stability compared to standard first-order Euler methods. By systematically analyzing the impact of different integration schemes, the authors show that their approach leads to superior semantic alignment and harmonic preservation in tasks like timbre transfer and genre modification. The framework consistently outperforms existing zero-shot editing methods in both objective metrics and perceptual audio quality, all while remaining flexible enough to handle arbitrary real-world recordings without additional training or optimization.