This paper introduces texture networks, a feed-forward approach for generating textures and stylized images. The method trains compact convolutional networks to generate multiple samples of a given texture and transfer artistic style from one image to another. Unlike previous methods that rely on slow optimization, texture networks use a learning stage to achieve faster and more memory-efficient generation. The networks are lightweight and can generate textures of high quality, comparable to Gatys et al.'s method, but significantly faster. The approach highlights the power of feed-forward models trained with complex loss functions. The networks can generate textures and process images of arbitrary size. The paper also presents extensive comparisons with other methods, showing that texture networks outperform previous approaches in both speed and quality. The method is particularly effective for texture synthesis and style transfer, and can be applied to a wide range of tasks, including video and mobile applications. The results demonstrate that texture networks can generate high-quality textures and stylized images, with perceptual quality comparable to optimization-based methods, but with significantly faster processing times. The paper also discusses the architecture of the networks and the training process, showing that the networks can be trained efficiently using a combination of texture and content loss functions. The results show that the method is robust and can handle a variety of styles and inputs. The paper concludes that texture networks offer a promising approach for texture synthesis and image stylization, with the potential for further improvements in stylization quality.This paper introduces texture networks, a feed-forward approach for generating textures and stylized images. The method trains compact convolutional networks to generate multiple samples of a given texture and transfer artistic style from one image to another. Unlike previous methods that rely on slow optimization, texture networks use a learning stage to achieve faster and more memory-efficient generation. The networks are lightweight and can generate textures of high quality, comparable to Gatys et al.'s method, but significantly faster. The approach highlights the power of feed-forward models trained with complex loss functions. The networks can generate textures and process images of arbitrary size. The paper also presents extensive comparisons with other methods, showing that texture networks outperform previous approaches in both speed and quality. The method is particularly effective for texture synthesis and style transfer, and can be applied to a wide range of tasks, including video and mobile applications. The results demonstrate that texture networks can generate high-quality textures and stylized images, with perceptual quality comparable to optimization-based methods, but with significantly faster processing times. The paper also discusses the architecture of the networks and the training process, showing that the networks can be trained efficiently using a combination of texture and content loss functions. The results show that the method is robust and can handle a variety of styles and inputs. The paper concludes that texture networks offer a promising approach for texture synthesis and image stylization, with the potential for further improvements in stylization quality.