The paper introduces a new class of loss functions called Deep Perceptual Similarity Metrics (DeepSiM) for image generation tasks. These loss functions are designed to mitigate the issue of over-smoothed results often encountered with traditional loss functions based on distances in the image space. Instead of computing distances directly in the image space, DeepSiM computes distances between image features extracted by deep neural networks, which better reflects perceptual similarity of images. The authors demonstrate three applications of DeepSiM: autoencoder training, a modification of a variational autoencoder, and inversion of deep convolutional networks. In all cases, the generated images look sharp and resemble natural images. The paper also discusses related work, including various neural network models for image generation and perceptual similarity metrics, and provides a detailed model architecture and training details. Experimental results show that DeepSiM outperforms traditional loss functions in terms of preserving fine details and realism in the generated images.The paper introduces a new class of loss functions called Deep Perceptual Similarity Metrics (DeepSiM) for image generation tasks. These loss functions are designed to mitigate the issue of over-smoothed results often encountered with traditional loss functions based on distances in the image space. Instead of computing distances directly in the image space, DeepSiM computes distances between image features extracted by deep neural networks, which better reflects perceptual similarity of images. The authors demonstrate three applications of DeepSiM: autoencoder training, a modification of a variational autoencoder, and inversion of deep convolutional networks. In all cases, the generated images look sharp and resemble natural images. The paper also discusses related work, including various neural network models for image generation and perceptual similarity metrics, and provides a detailed model architecture and training details. Experimental results show that DeepSiM outperforms traditional loss functions in terms of preserving fine details and realism in the generated images.