VITA-Group/DeblurGANv2

Employs a Feature Pyramid Network as the core generator component with a relativistic conditional GAN and double-scale discriminator for improved efficiency and quality. The architecture is backbone-agnostic, supporting plug-and-play substitution of feature extractors (Inception-ResNet-v2, MobileNet variants) to balance performance versus speed, enabling real-time video deblurring on lightweight backbones. Built in Python with TensorFlow/Keras, trained on standard benchmarks (GoPro, DVD, NFS) with pretrained models provided for inference via command-line prediction.

1,169 stars. No commits in the last 6 months.