about the contextual attention about generative_inpainting HOT 1 CLOSED

First thanks for your help on issue #38.

You can imagine that part of x is hallucinated (foreground), while another part of x is transformed from input image (background). We try to find if the hallucinated part of x have similar patches in background patch in space x, thus directly copy and paste to utilize long-range dependencies for better visual quality . The space x is learned by a neural network. One example is that if you mask one eye, most likely the filled result will be very similar to another eye in face. Other deep image inpainting systems usually fail in this case because they can not handle long-range dependencies spatially.

The attention module is with same merits with traditional image inpainting methods like PatchMatch: A randomized correspondence algorithm for structural image editing.

from generative_inpainting.