MEDIC: Zero-shot Music Editing with Disentangled Inversion Control

원문 초록

Text-guided diffusion models make a paradigm shift in audio generation, facilitating the adaptability of source audio to conform to specific textual prompts. Recent works introduce inversion techniques, like DDIM inversion, to zero-shot editing, exploiting pretrained diffusion models for audio modification. Nonetheless, our investigation exposes that DDIM inversion suffers from an accumulation of errors across each diffusion step, undermining its efficacy. Moreover, existing editing methods fail to achieve effective complex non-rigid music editing while maintaining essential content preservation and high editing fidelity. To counteract these issues, we introduce the Disentangled Inversion technique to disentangle the diffusion process into triple branches, rectifying the deviated path of the source branch caused by DDIM inversion. In addition, we propose the Harmonized Attention Control framework, which unifies the mutual self-attention control and cross-attention control with an intermediate Harmonic Branch to progressively achieve the desired harmonic and melodic information in the target music. Collectively, these innovations comprise the Disentangled Inversion Control (DIC) framework, enabling accurate music editing while safeguarding content integrity. To benchmark audio editing efficacy, we introduce ZoME-Bench, a comprehensive music editing benchmark hosting 1,100 samples spread across ten distinct editing categories. This facilitates both zero-shot and instruction-based music editing tasks. Our method achieves unparalleled performance in edit fidelity and essential content preservation, outperforming contemporary state-of-the-art inversion techniques.