DeblurGAN is an end-to-end learned method for motion deblurring using conditional adversarial networks. It achieves state-of-the-art performance in structural similarity and visual appearance. The method is 5 times faster than the closest competitor, DeepDeblur. DeblurGAN introduces a novel method for generating synthetic motion blurred images from sharp ones, enabling realistic dataset augmentation. The model, code, and dataset are available at https://github.com/KupynOrest/DeblurGAN. DeblurGAN treats deblurring as a special case of image-to-image translation. It uses a conditional generative adversarial network and a multi-component loss function. Unlike previous work, it uses Wasserstein GAN with gradient penalty and perceptual loss, which encourages solutions that are perceptually hard to distinguish from real sharp images and allows restoration of finer texture details. The method makes three contributions: (1) a loss and architecture that achieve state-of-the-art results in motion deblurring while being 5x faster than the fastest competitor; (2) a method based on random trajectories for generating a dataset for motion deblurring training in an automated fashion from sharp images; and (3) a novel dataset and method for evaluating deblurring algorithms based on how they improve object detection results. DeblurGAN is trained using a combination of adversarial and perceptual losses. The adversarial loss is based on WGAN-GP, while the perceptual loss is based on VGG-19 feature maps. The model is fully convolutional and can be applied to images of arbitrary size. It is trained on a single GPU and takes 6 days to train one network. DeblurGAN is evaluated on the GoPro and Kohler datasets, showing superior results in terms of structural similarity and visual appearance. It also improves object detection on blurred images, as demonstrated by YOLO results. DeblurGAN significantly outperforms competitors in terms of recall and F1 score. The method is able to handle blur caused by camera shake and object movement, and has more than 6x fewer parameters than Multi-scale CNN, which speeds up inference. The model is also able to generate realistic synthetic motion blur, allowing for a new benchmark and evaluation protocol based on object detection results.DeblurGAN is an end-to-end learned method for motion deblurring using conditional adversarial networks. It achieves state-of-the-art performance in structural similarity and visual appearance. The method is 5 times faster than the closest competitor, DeepDeblur. DeblurGAN introduces a novel method for generating synthetic motion blurred images from sharp ones, enabling realistic dataset augmentation. The model, code, and dataset are available at https://github.com/KupynOrest/DeblurGAN. DeblurGAN treats deblurring as a special case of image-to-image translation. It uses a conditional generative adversarial network and a multi-component loss function. Unlike previous work, it uses Wasserstein GAN with gradient penalty and perceptual loss, which encourages solutions that are perceptually hard to distinguish from real sharp images and allows restoration of finer texture details. The method makes three contributions: (1) a loss and architecture that achieve state-of-the-art results in motion deblurring while being 5x faster than the fastest competitor; (2) a method based on random trajectories for generating a dataset for motion deblurring training in an automated fashion from sharp images; and (3) a novel dataset and method for evaluating deblurring algorithms based on how they improve object detection results. DeblurGAN is trained using a combination of adversarial and perceptual losses. The adversarial loss is based on WGAN-GP, while the perceptual loss is based on VGG-19 feature maps. The model is fully convolutional and can be applied to images of arbitrary size. It is trained on a single GPU and takes 6 days to train one network. DeblurGAN is evaluated on the GoPro and Kohler datasets, showing superior results in terms of structural similarity and visual appearance. It also improves object detection on blurred images, as demonstrated by YOLO results. DeblurGAN significantly outperforms competitors in terms of recall and F1 score. The method is able to handle blur caused by camera shake and object movement, and has more than 6x fewer parameters than Multi-scale CNN, which speeds up inference. The model is also able to generate realistic synthetic motion blur, allowing for a new benchmark and evaluation protocol based on object detection results.