Real-Time Style Transfer and Super-Resolution with Perceptual Losses

By combining the approach of using pixel-wise loss between the output image and the ground truth image with the perception loss function based on high-level features extracted from pre-trained networks, we propose a forward network for image transformation tasks. Our work presents results for real-time style transfer and super-resolution, achieving similar qualitative results compared to methods based on optimization but with a much faster speed. By replacing pixel-wise loss with perception loss, we also achieve visually pleasing results for single-image super-resolution.