In recent years, generative models have made significant progress in the field of biology.Diffusion models have gradually become a research hotspot due to their advantages in generating high-quality samples and flexible generation mechanisms.Small interfering RNA (siRNA), as a powerful tool in molecular biology and therapeutics, achieves specific gene silencing through the RNA interference (RNAi) pathway. However, traditional siRNA sequence generation methods based on empirical rules and computational algorithms have limitations in accuracy and efficiency in exploring the sequence space.For example, there may be ambiguity when designing complex functional RNA sequences based on VAE.To better design siRNA molecules, this paper proposes a dual-branch collaborative diffusion model. This model reveals the complex interaction patterns between siRNA and mRNA, and implements intra-category contrastive learning for different therapeutic effects by constructing positive and negative sample pairs. In addition, the models use properties of siRNA to further optimize the generated siRNA sequences(such as thermodynamic),and we use Multi-Task Learning (MTL) to optimize multiple related tasks simultaneously.Experimental results show that the dual-diffusion model can generate entirely new sequences highly similar to natural siRNA, and these generated siRNA and their corresponding mRNA perform well in therapeutic effects. This not only proves the effectiveness of the model, but also provides new ideas and tools for RNAi drug design.