Old Photo Restoration with GANs

Generative model for restoring degraded old photos, plus image inpainting with the same architecture.

  • Pix2Pix GAN with PatchGAN discriminator: patch-level discrimination focused on the local detail lost in old photos; generator trained to produce realistic restorations
  • Image inpainting with the same architecture on a similar dataset
  • Dataset: Flickr-Faces-HQ Small (FFHQ-Small), 3,143 face images, 90:10 train/validation split
  • Restored outputs are sharper, more colorful, and less noisy than the originals; inference notebook included for trying the model

Code: github.com/bimarshak7/GAN-image-restore

Stack: Python, PyTorch, Jupyter