Recent advances in text-to-image diffusion models have enabled controllable image generation using spatial constraints such as bounding boxes. However, in interaction-centric generation tasks, existing methods such as InteractDiffusion still face critical challenges due to coupled spatial-semantic encoding in a unified feature space. This coupling often leads to modality interference, where spatial priors override text-driven action semantics, resulting in misaligned or unrealistic interactions. To address this problem, we introduce Decoupled Dual-Path Diffusion (DDP-Diffusion), a dual-path architecture that explicitly separates spatial control from semantic understanding. Specifically, we propose a novel MESA-Fusion module to improve spatial precision by transforming bounding box inputs into attention masks at the 16 × 16 layers of the U-Net. Meanwhile, the AALM module is designed to enhance semantic-physical alignment by encoding interaction semantics at the 32 × 32 and 64 × 64 layers. To better evaluate interaction generation, we propose a new metric, ActionScore, which combines text inversion with the multimodal Mini-CPM model and BERT-based similarity for fine-grained evaluation. Extensive experiments on the HICO-DET benchmark show that our model achieves state-of-the-art performance in both interaction quality (measured by ActionScore) and image fidelity (measured by FID and KID). Project page: https://github.com/BlackPuuuudding/DDP-Diffusion.git .