Recent progress in text-to-music generation has enabled high-quality audio synthesis from natural language prompts. However, such models are at risk of unintended replication, raising concerns regarding originality and intellectual property. While training-time mitigation strategies can address this issue, they typically require retraining or curated datasets, limiting their practicality for large-scale systems. Inference-time methods provide a lightweight alternative but often involve a trade-off between fidelity and memorization risk. This work introduces repulsive guidance (RG), a systematic inference-time mitigation strategy that reduces memorization without disrupting the intended conditional guidance from the text prompt. RG operates by enforcing divergence between dual diffusion trajectories through a repulsive term applied only during early denoising steps, without reversing the conditional guidance from the prompt. Experiments on MusicBench with the Tango model demonstrate that RG offers a complementary mitigation strategy, providing new insights into balancing fidelity and memorization risk.
Kim et al. (Thu,) studied this question.