context-encoder

Why we included this project

This repo is the reference implementation behind the CVPR 2016 context encoder paper, which showed that a GAN trained with both a reconstruction and an adversarial loss could learn useful visual features from nothing but the structure of missing image regions. The code bundles a semantic inpainting demo, training and testing routines for center and arbitrary random region inpainting, and pretrained models you can download. It is written in Torch and grew out of Soumith's DCGAN, so it reads easily if you are at home with the older deep learning stack. It is also research code last touched years ago, not a maintained product, so treat it as a way to learn how the ideas work rather than something to run in a real pipeline. Later inpainting and restoration tools still build on the same insights this paper introduced.