LAPTOP-Diff: Layer Pruning and Normalized Distillation for Compressing Diffusion Models This paper proposes LAPTOP-Diff, a method for compressing diffusion models by combining layer pruning and normalized distillation. The method aims to reduce the size of the U-Net in diffusion models while maintaining performance. The key contributions include: 1. Layer pruning: An automatic method for pruning layers in the U-Net, which is more efficient and scalable than previous handcrafted methods. The pruning criterion is based on the output loss, which is guaranteed to be additive, allowing for effective one-shot pruning. 2. Normalized feature distillation: A method to alleviate the imbalance issue in feature distillation during retraining. This is achieved by normalizing the feature loss terms based on the L2-norm of the teacher's feature maps. The method is evaluated on the SDXL and SDM-v1.5 models, achieving a minimal 4.0% decline in PickScore at a 50% pruning ratio, which is significantly better than other methods. The results show that the proposed method achieves the best performance across different models and pruning ratios. The method is also shown to be effective in reducing the model size while maintaining high-quality image generation. The paper also discusses the additivity property of the output loss criterion and its effectiveness in pruning. The method is validated through extensive experiments and ablation studies, demonstrating its effectiveness in compressing diffusion models.LAPTOP-Diff: Layer Pruning and Normalized Distillation for Compressing Diffusion Models This paper proposes LAPTOP-Diff, a method for compressing diffusion models by combining layer pruning and normalized distillation. The method aims to reduce the size of the U-Net in diffusion models while maintaining performance. The key contributions include: 1. Layer pruning: An automatic method for pruning layers in the U-Net, which is more efficient and scalable than previous handcrafted methods. The pruning criterion is based on the output loss, which is guaranteed to be additive, allowing for effective one-shot pruning. 2. Normalized feature distillation: A method to alleviate the imbalance issue in feature distillation during retraining. This is achieved by normalizing the feature loss terms based on the L2-norm of the teacher's feature maps. The method is evaluated on the SDXL and SDM-v1.5 models, achieving a minimal 4.0% decline in PickScore at a 50% pruning ratio, which is significantly better than other methods. The results show that the proposed method achieves the best performance across different models and pruning ratios. The method is also shown to be effective in reducing the model size while maintaining high-quality image generation. The paper also discusses the additivity property of the output loss criterion and its effectiveness in pruning. The method is validated through extensive experiments and ablation studies, demonstrating its effectiveness in compressing diffusion models.