DeblurGAN-v2 is a deep learning model for single image motion deblurring based on a relativistic conditional GAN with a double-scale discriminator. The generator uses a Feature Pyramid Network architecture that can flexibly work with different backbones ranging from heavy Inception-ResNet-v2 to lightweight MobileNet variants, enabling a trade-off between quality and speed. The model achieves competitive deblurring performance on benchmarks like GoPro and Kohler datasets, with the lightweight version running 10-100x faster than competitors, enabling potential real-time video deblurring.