Structural Pruning for Diffusion Models -- Supplementary Materials -- Gongfan Fang Xinyin Ma Xinchao Wang

In this document, we provide supplementary materials that we cannot fit into the main manuscript due to the page limit. It includes detailed explanations, visualization results, and several quantitative experiments. This section provides further insights into the coupled structures present in U-Net, which function as denoisers in diffusion models. In the context of structural pruning, it is crucial to prune layers with interdependencies simultaneously to avoid any potential structural issues [3]. To address these dependencies within U-Net, we leverage the use of DepGraph [1], which effectively handles most of the interdependencies.