Blind deblurring via Wasserstein guesswork

What it does

Feed it a blurry photo and it returns a sharper estimate. This is a PyTorch implementation of the 2017 DeblurGAN paper, built around a conditional Wasserstein GAN with gradient penalty and VGG-19 perceptual loss. It tackles blind motion deblurring, which means it recovers sharp images without prior knowledge of the blur kernel or camera motion.

The interesting bit

The authors note the same architecture transfers cleanly to other image-to-image translation problems—super-resolution, colorization, inpainting, dehazing—suggesting the network learns a general visual restoration prior rather than a blur-specific hack. The codebase borrows heavily from the pix2pix framework.

Key highlights

• Blind deblurring: no blur kernel or motion trajectory required
• Conditional Wasserstein GAN + gradient penalty + VGG-19 perceptual loss
• Inference requires only the generator weights
• Pre-trained weights and GoPRO evaluation data available via Google Drive
• Reported to generalize to super-resolution, dehazing, and colorization

Caveats

• CPU inference is explicitly untested; an NVIDIA GPU with CUDA CuDNN is effectively required

Verdict

A solid reference implementation for researchers working on image restoration or GAN-based translation. Look elsewhere if you need CPU-only inference or a maintained, general-purpose deblurring tool.