This paper introduces a method for image transformation tasks that combines feed-forward convolutional neural networks with perceptual loss functions. The approach uses high-level features extracted from pretrained networks to train networks for tasks such as image style transfer and single-image super-resolution. The method achieves similar qualitative results to existing optimization-based methods but is significantly faster, up to three orders of magnitude faster for style transfer. For super-resolution, replacing the per-pixel loss with a perceptual loss produces visually pleasing results. The paper presents two main tasks: style transfer and single-image super-resolution. For style transfer, the goal is to generate an image that combines the content of a target content image with the style of a target style image. The method uses a perceptual loss function based on high-level features from a pretrained network, allowing the model to learn semantic knowledge from the loss network. For super-resolution, the method uses a perceptual loss function to better reconstruct fine details compared to per-pixel loss methods. The paper also discusses related work, including feed-forward image transformation and perceptual optimization. It describes the architecture of the image transformation networks, which include residual blocks and use a pretrained loss network to define perceptual loss functions. The method is evaluated on two tasks, showing that it achieves comparable performance to existing methods but with significantly faster speeds. The results demonstrate that the perceptual loss function allows the model to better reconstruct fine details and edges, leading to visually pleasing results. The paper concludes that combining feed-forward networks with perceptual loss functions offers a promising approach for image transformation tasks.This paper introduces a method for image transformation tasks that combines feed-forward convolutional neural networks with perceptual loss functions. The approach uses high-level features extracted from pretrained networks to train networks for tasks such as image style transfer and single-image super-resolution. The method achieves similar qualitative results to existing optimization-based methods but is significantly faster, up to three orders of magnitude faster for style transfer. For super-resolution, replacing the per-pixel loss with a perceptual loss produces visually pleasing results. The paper presents two main tasks: style transfer and single-image super-resolution. For style transfer, the goal is to generate an image that combines the content of a target content image with the style of a target style image. The method uses a perceptual loss function based on high-level features from a pretrained network, allowing the model to learn semantic knowledge from the loss network. For super-resolution, the method uses a perceptual loss function to better reconstruct fine details compared to per-pixel loss methods. The paper also discusses related work, including feed-forward image transformation and perceptual optimization. It describes the architecture of the image transformation networks, which include residual blocks and use a pretrained loss network to define perceptual loss functions. The method is evaluated on two tasks, showing that it achieves comparable performance to existing methods but with significantly faster speeds. The results demonstrate that the perceptual loss function allows the model to better reconstruct fine details and edges, leading to visually pleasing results. The paper concludes that combining feed-forward networks with perceptual loss functions offers a promising approach for image transformation tasks.