This is a demo for the 2019 ICCV paper: "Seeing what a GAN cannot generate" and it allows you to see how well a Generative Adversarial Network (GAN) trained on church images can reconstruct an input image. The demo highlights what the GAN misses in that individual image -- what it "GAN not see..."

This is a joint project of MIT CSAIL and the MIT-IBM Watson AI Lab by Bau, Zhu, Wulff, Peebles, Strobelt, Zhou, and Torralba.

Choose one of the examples. Images with white border are produced by a GAN. Black border indicates real-world images from the training set:

to project main page

We collect information about your activity. Since we care about your privacy, all these information are fully anonymized. Your uploaded image will not be stored and is only used to calculate the results. Check the links for more information: Terms of Use and Privacy