DeblurGAN

Why we included this project

Motion blur quietly ruins otherwise usable frames from phone cameras, old film scans, and surveillance footage. DeblurGAN is a PyTorch implementation of a GAN that takes a blurry image and returns a sharp estimate in a single pass, pairing a conditional Wasserstein GAN with a VGG-19-based perceptual loss. The authors show the same architecture transfers to other image-to-image jobs like super-resolution and colorization. Pretrained generator weights and a straightforward test script let you run it on your own images right away, and the repo includes the pairing and training scripts if you want to retrain on a domain closer to your own footage. With an NVIDIA GPU and CUDA, it is a quick way to see how adversarial training plus perceptual losses hold up on real blur.