The paper introduces Latent Dataset Distillation with Diffusion Models (LD3M), a novel method that combines diffusion models with dataset distillation to generate synthetic samples from large datasets. LD3M addresses the challenges of storing large datasets and the presence of non-influential samples by condensing the dataset into a smaller set of synthetic images. The key contributions of LD3M include: 1. **Diffusion Models for Latent Space**: LD3M leverages diffusion models to optimize latent codes in the latent space, rather than directly optimizing pixel values. This approach improves gradient flow and generalizes better across different architectures and high-resolution images. 2. **Enhanced Gradient Flow**: By adjusting the number of diffusion steps, LD3M allows for a trade-off between distillation speed and dataset quality. The method significantly outperforms state-of-the-art methods by up to 4.8 p.p. and 4.2 p.p. for 1 and 10 images per class, respectively, on several ImageNet subsets and high resolutions (128x128 and 256x256). 3. **Flexibility and Adaptability**: LD3M is compatible with any distillation algorithm and can be easily extended to new diffusion models. It also offers a straightforward initialization of the distilled dataset using a pre-trained encoder, which is more efficient than GAN-based methods like GLaD. 4. **Cross-Architecture Performance**: LD3M consistently outperforms GLaD in cross-architecture evaluations, demonstrating superior performance on unseen architectures for both 1 and 10 images per class. 5. **High-Resolution Image Generation**: LD3M generates high-quality synthetic images at resolutions of 256x256, achieving better results than other methods in this domain. The paper evaluates LD3M on various datasets and architectures, showing its effectiveness and flexibility. LD3M not only improves the quality of synthetic datasets but also enhances the overall accuracy of trained models, making it a significant advancement in the field of dataset distillation.The paper introduces Latent Dataset Distillation with Diffusion Models (LD3M), a novel method that combines diffusion models with dataset distillation to generate synthetic samples from large datasets. LD3M addresses the challenges of storing large datasets and the presence of non-influential samples by condensing the dataset into a smaller set of synthetic images. The key contributions of LD3M include: 1. **Diffusion Models for Latent Space**: LD3M leverages diffusion models to optimize latent codes in the latent space, rather than directly optimizing pixel values. This approach improves gradient flow and generalizes better across different architectures and high-resolution images. 2. **Enhanced Gradient Flow**: By adjusting the number of diffusion steps, LD3M allows for a trade-off between distillation speed and dataset quality. The method significantly outperforms state-of-the-art methods by up to 4.8 p.p. and 4.2 p.p. for 1 and 10 images per class, respectively, on several ImageNet subsets and high resolutions (128x128 and 256x256). 3. **Flexibility and Adaptability**: LD3M is compatible with any distillation algorithm and can be easily extended to new diffusion models. It also offers a straightforward initialization of the distilled dataset using a pre-trained encoder, which is more efficient than GAN-based methods like GLaD. 4. **Cross-Architecture Performance**: LD3M consistently outperforms GLaD in cross-architecture evaluations, demonstrating superior performance on unseen architectures for both 1 and 10 images per class. 5. **High-Resolution Image Generation**: LD3M generates high-quality synthetic images at resolutions of 256x256, achieving better results than other methods in this domain. The paper evaluates LD3M on various datasets and architectures, showing its effectiveness and flexibility. LD3M not only improves the quality of synthetic datasets but also enhances the overall accuracy of trained models, making it a significant advancement in the field of dataset distillation.