Multi-Code Manifold Fusion for Generative Dataset Distillation

The most important aim of dataset distillation is to compress a bulky dataset to a small artificial dataset and preserve a less loss in training performance than using the original dataset. The parametric method should have a better performance with high-resolution datasets by converting the optimization task in the pixel space into the feature space of the generative model. Nonetheless, feature-based optimization has a direct effect on the inherent consistency of the generative model manifold, subsequently restricting its crossarchitecture generalization of synthetic data. We introduce the generative framework Multicode Manifold Fusion (MCMF), which is more restrictive on the prior conditions in the data generation process since it strictly restricts the latent space behaviour of the frozen generative model. This involves a new subspace parameterization scheme that is aimed at contributing to the expressive power of low-dimensional latent code. This strategy constructs the representation subspace by leveraging the semantic anchors obtained through clustering and employs a learnable channel attention mechanism to optimize the interpolation operation. This strategy both enhances expressive power and diversity while adhering to strict prior constraints. We conduct comprehensive experiments on datasets of different resolutions to verify the outstanding performance of the MCMF framework. It is worth noting that this method possesses excellent cross-architecture transferability, which fully demonstrates that the framework we proposed can effectively learn the essential semantic information of the data.