DeblurGAN-v2 is a new end-to-end generative adversarial network (GAN) for single image motion deblurring that significantly improves deblurring efficiency, quality, and flexibility. It is based on a relativistic conditional GAN with a double-scale discriminator and introduces the Feature Pyramid Network (FPN) as a core component in the generator. The FPN allows the model to work with various backbones, balancing performance and efficiency. Sophisticated backbones like Inception-ResNet-v2 achieve state-of-the-art results, while lightweight backbones like MobileNet and MobileNet-DSC enable 10-100 times faster inference, making real-time video deblurring possible. DeblurGAN-v2 achieves competitive performance on multiple benchmarks, demonstrating high quality and efficiency. It is also effective for general image restoration tasks. The model is implemented with various backbones, allowing flexibility in performance and efficiency trade-offs. DeblurGAN-v2 outperforms existing methods in terms of PSNR, SSIM, and perceptual quality, while being significantly faster. It is also effective for video deblurring and general image restoration. The model is available at https://github.com/KupynOrest/DeblurGANv2.DeblurGAN-v2 is a new end-to-end generative adversarial network (GAN) for single image motion deblurring that significantly improves deblurring efficiency, quality, and flexibility. It is based on a relativistic conditional GAN with a double-scale discriminator and introduces the Feature Pyramid Network (FPN) as a core component in the generator. The FPN allows the model to work with various backbones, balancing performance and efficiency. Sophisticated backbones like Inception-ResNet-v2 achieve state-of-the-art results, while lightweight backbones like MobileNet and MobileNet-DSC enable 10-100 times faster inference, making real-time video deblurring possible. DeblurGAN-v2 achieves competitive performance on multiple benchmarks, demonstrating high quality and efficiency. It is also effective for general image restoration tasks. The model is implemented with various backbones, allowing flexibility in performance and efficiency trade-offs. DeblurGAN-v2 outperforms existing methods in terms of PSNR, SSIM, and perceptual quality, while being significantly faster. It is also effective for video deblurring and general image restoration. The model is available at https://github.com/KupynOrest/DeblurGANv2.