IIDiff: Learning Cross-Domain Frequency Transitions with Diffusion Mixture-of-Experts

Cross-domain time-series generation is a challenging task for diffusion models due to the strong requirement for domain adaptability. While frequency-based Mixture-of-Experts (MoE) methods have shown promise in domain adaptation, their direct integration with diffusion models faces critical issues from intra- and inter-domain. For intra-domain aspect, the absence of frequency transition compliance leads to imbalance pattern learning across denoising steps and degrade model performance. For the inter-domain aspect, the domain-specific transition variation bring difficulties for domain adaptation. To address these challenges, we propose IIDiff, a frequency-based diffusion MoE framework equipped with progressive frequency specializing and dynamic transition matching. The progressive specializing allows experts dynamically adjust frequency receptive fields to capture underlying patterns aligning with the intra-domain transition, enhancing the generation fidelity. The transition matching leverages the frequency noise distribution property to analysis the instantaneous transition process, enabling adaptive routing with domain-specific transition. Extensive experiments on 12 real-world datasets across 4 domains demonstrate our model's superiority. Compared with SOTA baselines, our model achieves an average 27.59\% improvement in KL divergence and 6.00\% improvement in MMD, confirming its robust domain adaptability and ability to generate high-fidelity cross-domain time-series.