A world model allows robots to understand and predict the interplay between their actions and environmental dynamics. Recent advancements in diffusion models have significantly improved the quality of image frame generation in simulated environments, contributing to the development of more robust and generalized world models. However, these diffusion-based world models often depend on discrete inputs, such as keyboard commands, which limit their applicability to continuous real-world robotic control. To address this limitation, we propose a novel framework that integrates contrastive learning to align visual and proprioceptive modalities (e.g., joint positions) within a shared latent space. This shared latent space facilitates accurate cross-modal predictions between visual scenes and proprioceptive states. By combining this latent representation with a diffusion model, our world model can generate long-term future visual scenes by leveraging both initial visual observations and proprioceptive states. Experimental results demonstrate that the proposed framework generates high-fidelity, long-term future visual scenes when provided with target proprioceptive data. This capability enables robots to plan their motions solely based on the generated images, enabling imagination-based planning.