Open SURABHI-GUPTA opened 3 years ago
The point of this paper was to show how far you could get using only the prior implicit in the model architecture itself. Since the usual convolutional model doesn't have any built-in notions of noses, eyes etc. so it's not at all competitive to a trained inpainting network... But you can always try to get an idea of how well (or badly) it does on your model of choice, what I especially liked about this paper was how dead simple it was to implement.
@DmitryUlyanov @albarji @cclauss How about impainting on face images? Let's say a portion of the face is occluded. Does your impainting code work well on such images?