This paper proposes LD3M, a novel framework for latent dataset distillation with generative priors that improves the gradient flow of diffusion models during the distillation process.

We show the effects of vanishing gradients on GNNs.