The paper introduces D4M (Dataset Distillation via Disentangled Diffusion Model), an efficient and economical framework for dataset distillation. Unlike traditional methods that rely on bi-level optimization and specific matching architectures, D4M employs a latent diffusion model to ensure consistency between the real and synthetic image spaces, enhancing cross-architecture generalization. The method integrates label information into category prototypes, making the distilled datasets versatile and adaptable to various architectures. Through extensive experiments, D4M demonstrates superior performance and robust generalization, outperforming state-of-the-art methods across multiple datasets, including ImageNet-1K and CIFAR-10/100. Key contributions include the introduction of Training-Time Matching (TTM) to improve distillation efficiency and the integration of multi-modal fusion embedding to enhance model performance. The method also addresses the limitations of existing DD frameworks, such as computational costs and architectural dependency, by leveraging the architecture-free nature of diffusion models.The paper introduces D4M (Dataset Distillation via Disentangled Diffusion Model), an efficient and economical framework for dataset distillation. Unlike traditional methods that rely on bi-level optimization and specific matching architectures, D4M employs a latent diffusion model to ensure consistency between the real and synthetic image spaces, enhancing cross-architecture generalization. The method integrates label information into category prototypes, making the distilled datasets versatile and adaptable to various architectures. Through extensive experiments, D4M demonstrates superior performance and robust generalization, outperforming state-of-the-art methods across multiple datasets, including ImageNet-1K and CIFAR-10/100. Key contributions include the introduction of Training-Time Matching (TTM) to improve distillation efficiency and the integration of multi-modal fusion embedding to enhance model performance. The method also addresses the limitations of existing DD frameworks, such as computational costs and architectural dependency, by leveraging the architecture-free nature of diffusion models.