DesignEdit is a unified framework integrating various spatially aware image editing functionalities. It achieves this by decomposing spatial-aware image editing tasks into two subtasks: decomposition and fusion of latent representations at multiple levels. Firstly, the latent representation of the source image is divided into multiple layers, including several object layers and an incomplete background layer that requires reliable repair. To avoid additional tuning, we further explore the repair capabilities within the self-attention mechanism, introducing a key-masking self-attention scheme that can propagate contextual information from surrounding areas to the masked areas while reducing the influence of areas outside the mask. Secondly, we propose an instruction-based latent fusion method to paste the multi-level latent representations onto the canvas latent space. We also introduce a latent space artifact suppression mechanism to enhance the repair quality. Due to the inherent modular advantages of this multi-level representation, we can achieve precise image editing, and our method has shown excellent performance on multiple editing tasks, surpassing the latest spatial editing methods.