LD-Pruner is a novel structured pruning method for compressing Latent Diffusion Models (LDMs) while preserving performance. Traditional pruning methods are not tailored to LDMs, which have unique characteristics such as high computational costs and lack of task-agnostic evaluation methods. LD-Pruner addresses these challenges by leveraging the latent space during pruning, enabling the quantification of pruning impact without task dependency. This approach allows for efficient pruning of components with minimal impact on output, leading to faster convergence during training and reduced parameter count without significant performance degradation. The method was tested on three tasks: text-to-image (T2I), unconditional image generation (UIG), and unconditional audio generation (UAG). On the Stable Diffusion model, LD-Pruner reduced inference time by 34.9% while improving FID by 5.2% on the MS-COCO T2I benchmark. The approach achieves a compressed model with improved inference speed and reduced parameters, maintaining minimal performance loss. The method uses a task-agnostic metric to evaluate operator importance, enabling structured pruning based on latent space analysis. This leads to efficient pruning of less impactful components, resulting in a compact model with high performance. The method is applicable across various tasks and offers a trade-off between speed and performance, with a threshold-like behavior in performance degradation as compression increases. The approach is suitable for LDMs where most computational costs occur in the U-Net, and it does not account for operator dependencies, which could lead to suboptimal pruning decisions. The method is task-agnostic, leveraging latent space to guide pruning, and provides a robust, quantifiable measure for pruning decisions. The results demonstrate the effectiveness of LD-Pruner in compressing LDMs while maintaining performance, making them more applicable in resource-constrained environments.LD-Pruner is a novel structured pruning method for compressing Latent Diffusion Models (LDMs) while preserving performance. Traditional pruning methods are not tailored to LDMs, which have unique characteristics such as high computational costs and lack of task-agnostic evaluation methods. LD-Pruner addresses these challenges by leveraging the latent space during pruning, enabling the quantification of pruning impact without task dependency. This approach allows for efficient pruning of components with minimal impact on output, leading to faster convergence during training and reduced parameter count without significant performance degradation. The method was tested on three tasks: text-to-image (T2I), unconditional image generation (UIG), and unconditional audio generation (UAG). On the Stable Diffusion model, LD-Pruner reduced inference time by 34.9% while improving FID by 5.2% on the MS-COCO T2I benchmark. The approach achieves a compressed model with improved inference speed and reduced parameters, maintaining minimal performance loss. The method uses a task-agnostic metric to evaluate operator importance, enabling structured pruning based on latent space analysis. This leads to efficient pruning of less impactful components, resulting in a compact model with high performance. The method is applicable across various tasks and offers a trade-off between speed and performance, with a threshold-like behavior in performance degradation as compression increases. The approach is suitable for LDMs where most computational costs occur in the U-Net, and it does not account for operator dependencies, which could lead to suboptimal pruning decisions. The method is task-agnostic, leveraging latent space to guide pruning, and provides a robust, quantifiable measure for pruning decisions. The results demonstrate the effectiveness of LD-Pruner in compressing LDMs while maintaining performance, making them more applicable in resource-constrained environments.